8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "precompiled.hpp"
26 #include "ci/ciUtilities.hpp"
27 #include "compiler/compileLog.hpp"
28 #include "gc/shared/barrierSet.hpp"
29 #include "gc/shared/c2/barrierSetC2.hpp"
30 #include "interpreter/interpreter.hpp"
31 #include "memory/resourceArea.hpp"
32 #include "opto/addnode.hpp"
33 #include "opto/castnode.hpp"
34 #include "opto/convertnode.hpp"
35 #include "opto/graphKit.hpp"
36 #include "opto/idealKit.hpp"
37 #include "opto/intrinsicnode.hpp"
38 #include "opto/locknode.hpp"
39 #include "opto/machnode.hpp"
40 #include "opto/opaquenode.hpp"
41 #include "opto/parse.hpp"
42 #include "opto/rootnode.hpp"
43 #include "opto/runtime.hpp"
44 #include "runtime/deoptimization.hpp"
45 #include "runtime/sharedRuntime.hpp"
46
47 //----------------------------GraphKit-----------------------------------------
48 // Main utility constructor.
49 GraphKit::GraphKit(JVMState* jvms)
50 : Phase(Phase::Parser),
51 _env(C->env()),
52 _gvn(*C->initial_gvn()),
53 _barrier_set(BarrierSet::barrier_set()->barrier_set_c2())
54 {
55 _exceptions = jvms->map()->next_exception();
56 if (_exceptions != NULL) jvms->map()->set_next_exception(NULL);
57 set_jvms(jvms);
58 }
59
60 // Private constructor for parser.
61 GraphKit::GraphKit()
62 : Phase(Phase::Parser),
63 _env(C->env()),
64 _gvn(*C->initial_gvn()),
65 _barrier_set(BarrierSet::barrier_set()->barrier_set_c2())
66 {
67 _exceptions = NULL;
68 set_map(NULL);
69 debug_only(_sp = -99);
70 debug_only(set_bci(-99));
71 }
72
73
74
75 //---------------------------clean_stack---------------------------------------
76 // Clear away rubbish from the stack area of the JVM state.
77 // This destroys any arguments that may be waiting on the stack.
806 tty->print_cr("Zombie local %d: ", local);
807 jvms->dump();
808 }
809 return false;
810 }
811 }
812 }
813 return true;
814 }
815
816 #endif //ASSERT
817
818 // Helper function for enforcing certain bytecodes to reexecute if
819 // deoptimization happens
820 static bool should_reexecute_implied_by_bytecode(JVMState *jvms, bool is_anewarray) {
821 ciMethod* cur_method = jvms->method();
822 int cur_bci = jvms->bci();
823 if (cur_method != NULL && cur_bci != InvocationEntryBci) {
824 Bytecodes::Code code = cur_method->java_code_at_bci(cur_bci);
825 return Interpreter::bytecode_should_reexecute(code) ||
826 (is_anewarray && code == Bytecodes::_multianewarray);
827 // Reexecute _multianewarray bytecode which was replaced with
828 // sequence of [a]newarray. See Parse::do_multianewarray().
829 //
830 // Note: interpreter should not have it set since this optimization
831 // is limited by dimensions and guarded by flag so in some cases
832 // multianewarray() runtime calls will be generated and
833 // the bytecode should not be reexecutes (stack will not be reset).
834 } else
835 return false;
836 }
837
838 // Helper function for adding JVMState and debug information to node
839 void GraphKit::add_safepoint_edges(SafePointNode* call, bool must_throw) {
840 // Add the safepoint edges to the call (or other safepoint).
841
842 // Make sure dead locals are set to top. This
843 // should help register allocation time and cut down on the size
844 // of the deoptimization information.
845 assert(dead_locals_are_killed(), "garbage in debug info before safepoint");
846
847 // Walk the inline list to fill in the correct set of JVMState's
848 // Also fill in the associated edges for each JVMState.
849
850 // If the bytecode needs to be reexecuted we need to put
851 // the arguments back on the stack.
852 const bool should_reexecute = jvms()->should_reexecute();
853 JVMState* youngest_jvms = should_reexecute ? sync_jvms_for_reexecute() : sync_jvms();
854
855 // NOTE: set_bci (called from sync_jvms) might reset the reexecute bit to
1059 ciSignature* declared_signature = NULL;
1060 ciMethod* ignored_callee = method()->get_method_at_bci(bci(), ignored_will_link, &declared_signature);
1061 assert(declared_signature != NULL, "cannot be null");
1062 inputs = declared_signature->arg_size_for_bc(code);
1063 int size = declared_signature->return_type()->size();
1064 depth = size - inputs;
1065 }
1066 break;
1067
1068 case Bytecodes::_multianewarray:
1069 {
1070 ciBytecodeStream iter(method());
1071 iter.reset_to_bci(bci());
1072 iter.next();
1073 inputs = iter.get_dimensions();
1074 assert(rsize == 1, "");
1075 depth = rsize - inputs;
1076 }
1077 break;
1078
1079 case Bytecodes::_ireturn:
1080 case Bytecodes::_lreturn:
1081 case Bytecodes::_freturn:
1082 case Bytecodes::_dreturn:
1083 case Bytecodes::_areturn:
1084 assert(rsize == -depth, "");
1085 inputs = rsize;
1086 break;
1087
1088 case Bytecodes::_jsr:
1089 case Bytecodes::_jsr_w:
1090 inputs = 0;
1091 depth = 1; // S.B. depth=1, not zero
1092 break;
1093
1094 default:
1095 // bytecode produces a typed result
1096 inputs = rsize - depth;
1097 assert(inputs >= 0, "");
1098 break;
1184 // the incoming address with NULL casted away. You are allowed to use the
1185 // not-null value only if you are control dependent on the test.
1186 #ifndef PRODUCT
1187 extern int explicit_null_checks_inserted,
1188 explicit_null_checks_elided;
1189 #endif
1190 Node* GraphKit::null_check_common(Node* value, BasicType type,
1191 // optional arguments for variations:
1192 bool assert_null,
1193 Node* *null_control,
1194 bool speculative) {
1195 assert(!assert_null || null_control == NULL, "not both at once");
1196 if (stopped()) return top();
1197 NOT_PRODUCT(explicit_null_checks_inserted++);
1198
1199 // Construct NULL check
1200 Node *chk = NULL;
1201 switch(type) {
1202 case T_LONG : chk = new CmpLNode(value, _gvn.zerocon(T_LONG)); break;
1203 case T_INT : chk = new CmpINode(value, _gvn.intcon(0)); break;
1204 case T_ARRAY : // fall through
1205 type = T_OBJECT; // simplify further tests
1206 case T_OBJECT : {
1207 const Type *t = _gvn.type( value );
1208
1209 const TypeOopPtr* tp = t->isa_oopptr();
1210 if (tp != NULL && tp->klass() != NULL && !tp->klass()->is_loaded()
1211 // Only for do_null_check, not any of its siblings:
1212 && !assert_null && null_control == NULL) {
1213 // Usually, any field access or invocation on an unloaded oop type
1214 // will simply fail to link, since the statically linked class is
1215 // likely also to be unloaded. However, in -Xcomp mode, sometimes
1216 // the static class is loaded but the sharper oop type is not.
1217 // Rather than checking for this obscure case in lots of places,
1218 // we simply observe that a null check on an unloaded class
1219 // will always be followed by a nonsense operation, so we
1220 // can just issue the uncommon trap here.
1221 // Our access to the unloaded class will only be correct
1222 // after it has been loaded and initialized, which requires
1223 // a trip through the interpreter.
1355 }
1356
1357 if (assert_null) {
1358 // Cast obj to null on this path.
1359 replace_in_map(value, zerocon(type));
1360 return zerocon(type);
1361 }
1362
1363 // Cast obj to not-null on this path, if there is no null_control.
1364 // (If there is a null_control, a non-null value may come back to haunt us.)
1365 if (type == T_OBJECT) {
1366 Node* cast = cast_not_null(value, false);
1367 if (null_control == NULL || (*null_control) == top())
1368 replace_in_map(value, cast);
1369 value = cast;
1370 }
1371
1372 return value;
1373 }
1374
1375
1376 //------------------------------cast_not_null----------------------------------
1377 // Cast obj to not-null on this path
1378 Node* GraphKit::cast_not_null(Node* obj, bool do_replace_in_map) {
1379 const Type *t = _gvn.type(obj);
1380 const Type *t_not_null = t->join_speculative(TypePtr::NOTNULL);
1381 // Object is already not-null?
1382 if( t == t_not_null ) return obj;
1383
1384 Node *cast = new CastPPNode(obj,t_not_null);
1385 cast->init_req(0, control());
1386 cast = _gvn.transform( cast );
1387
1388 // Scan for instances of 'obj' in the current JVM mapping.
1389 // These instances are known to be not-null after the test.
1390 if (do_replace_in_map)
1391 replace_in_map(obj, cast);
1392
1393 return cast; // Return casted value
1394 }
1395
1396 // Sometimes in intrinsics, we implicitly know an object is not null
1397 // (there's no actual null check) so we can cast it to not null. In
1398 // the course of optimizations, the input to the cast can become null.
1487 int adr_idx,
1488 MemNode::MemOrd mo,
1489 LoadNode::ControlDependency control_dependency,
1490 bool require_atomic_access,
1491 bool unaligned,
1492 bool mismatched,
1493 bool unsafe) {
1494 assert(adr_idx != Compile::AliasIdxTop, "use other make_load factory" );
1495 const TypePtr* adr_type = NULL; // debug-mode-only argument
1496 debug_only(adr_type = C->get_adr_type(adr_idx));
1497 Node* mem = memory(adr_idx);
1498 Node* ld;
1499 if (require_atomic_access && bt == T_LONG) {
1500 ld = LoadLNode::make_atomic(ctl, mem, adr, adr_type, t, mo, control_dependency, unaligned, mismatched, unsafe);
1501 } else if (require_atomic_access && bt == T_DOUBLE) {
1502 ld = LoadDNode::make_atomic(ctl, mem, adr, adr_type, t, mo, control_dependency, unaligned, mismatched, unsafe);
1503 } else {
1504 ld = LoadNode::make(_gvn, ctl, mem, adr, adr_type, t, bt, mo, control_dependency, unaligned, mismatched, unsafe);
1505 }
1506 ld = _gvn.transform(ld);
1507 if (((bt == T_OBJECT) && C->do_escape_analysis()) || C->eliminate_boxing()) {
1508 // Improve graph before escape analysis and boxing elimination.
1509 record_for_igvn(ld);
1510 }
1511 return ld;
1512 }
1513
1514 Node* GraphKit::store_to_memory(Node* ctl, Node* adr, Node *val, BasicType bt,
1515 int adr_idx,
1516 MemNode::MemOrd mo,
1517 bool require_atomic_access,
1518 bool unaligned,
1519 bool mismatched,
1520 bool unsafe) {
1521 assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory" );
1522 const TypePtr* adr_type = NULL;
1523 debug_only(adr_type = C->get_adr_type(adr_idx));
1524 Node *mem = memory(adr_idx);
1525 Node* st;
1526 if (require_atomic_access && bt == T_LONG) {
1527 st = StoreLNode::make_atomic(ctl, mem, adr, adr_type, val, mo);
1538 }
1539 if (unsafe) {
1540 st->as_Store()->set_unsafe_access();
1541 }
1542 st = _gvn.transform(st);
1543 set_memory(st, adr_idx);
1544 // Back-to-back stores can only remove intermediate store with DU info
1545 // so push on worklist for optimizer.
1546 if (mem->req() > MemNode::Address && adr == mem->in(MemNode::Address))
1547 record_for_igvn(st);
1548
1549 return st;
1550 }
1551
1552 Node* GraphKit::access_store_at(Node* obj,
1553 Node* adr,
1554 const TypePtr* adr_type,
1555 Node* val,
1556 const Type* val_type,
1557 BasicType bt,
1558 DecoratorSet decorators) {
1559 // Transformation of a value which could be NULL pointer (CastPP #NULL)
1560 // could be delayed during Parse (for example, in adjust_map_after_if()).
1561 // Execute transformation here to avoid barrier generation in such case.
1562 if (_gvn.type(val) == TypePtr::NULL_PTR) {
1563 val = _gvn.makecon(TypePtr::NULL_PTR);
1564 }
1565
1566 if (stopped()) {
1567 return top(); // Dead path ?
1568 }
1569
1570 assert(val != NULL, "not dead path");
1571
1572 C2AccessValuePtr addr(adr, adr_type);
1573 C2AccessValue value(val, val_type);
1574 C2ParseAccess access(this, decorators | C2_WRITE_ACCESS, bt, obj, addr);
1575 if (access.is_raw()) {
1576 return _barrier_set->BarrierSetC2::store_at(access, value);
1577 } else {
1578 return _barrier_set->store_at(access, value);
1579 }
1580 }
1581
1582 Node* GraphKit::access_load_at(Node* obj, // containing obj
1583 Node* adr, // actual adress to store val at
1584 const TypePtr* adr_type,
1585 const Type* val_type,
1586 BasicType bt,
1587 DecoratorSet decorators) {
1588 if (stopped()) {
1589 return top(); // Dead path ?
1590 }
1591
1592 C2AccessValuePtr addr(adr, adr_type);
1593 C2ParseAccess access(this, decorators | C2_READ_ACCESS, bt, obj, addr);
1594 if (access.is_raw()) {
1595 return _barrier_set->BarrierSetC2::load_at(access, val_type);
1596 } else {
1597 return _barrier_set->load_at(access, val_type);
1598 }
1599 }
1600
1601 Node* GraphKit::access_load(Node* adr, // actual adress to load val at
1602 const Type* val_type,
1603 BasicType bt,
1604 DecoratorSet decorators) {
1605 if (stopped()) {
1606 return top(); // Dead path ?
1607 }
1608
1609 C2AccessValuePtr addr(adr, NULL);
1610 C2ParseAccess access(this, decorators | C2_READ_ACCESS, bt, NULL, addr);
1611 if (access.is_raw()) {
1612 return _barrier_set->BarrierSetC2::load_at(access, val_type);
1613 } else {
1671 }
1672 }
1673
1674 Node* GraphKit::access_atomic_add_at(Node* obj,
1675 Node* adr,
1676 const TypePtr* adr_type,
1677 int alias_idx,
1678 Node* new_val,
1679 const Type* value_type,
1680 BasicType bt,
1681 DecoratorSet decorators) {
1682 C2AccessValuePtr addr(adr, adr_type);
1683 C2AtomicParseAccess access(this, decorators | C2_READ_ACCESS | C2_WRITE_ACCESS, bt, obj, addr, alias_idx);
1684 if (access.is_raw()) {
1685 return _barrier_set->BarrierSetC2::atomic_add_at(access, new_val, value_type);
1686 } else {
1687 return _barrier_set->atomic_add_at(access, new_val, value_type);
1688 }
1689 }
1690
1691 void GraphKit::access_clone(Node* src, Node* dst, Node* size, bool is_array) {
1692 return _barrier_set->clone(this, src, dst, size, is_array);
1693 }
1694
1695 Node* GraphKit::access_resolve(Node* n, DecoratorSet decorators) {
1696 // Use stronger ACCESS_WRITE|ACCESS_READ by default.
1697 if ((decorators & (ACCESS_READ | ACCESS_WRITE)) == 0) {
1698 decorators |= ACCESS_READ | ACCESS_WRITE;
1699 }
1700 return _barrier_set->resolve(this, n, decorators);
1701 }
1702
1703 //-------------------------array_element_address-------------------------
1704 Node* GraphKit::array_element_address(Node* ary, Node* idx, BasicType elembt,
1705 const TypeInt* sizetype, Node* ctrl) {
1706 uint shift = exact_log2(type2aelembytes(elembt));
1707 uint header = arrayOopDesc::base_offset_in_bytes(elembt);
1708
1709 // short-circuit a common case (saves lots of confusing waste motion)
1710 jint idx_con = find_int_con(idx, -1);
1711 if (idx_con >= 0) {
1712 intptr_t offset = header + ((intptr_t)idx_con << shift);
1713 return basic_plus_adr(ary, offset);
1714 }
1715
1716 // must be correct type for alignment purposes
1717 Node* base = basic_plus_adr(ary, header);
1718 idx = Compile::conv_I2X_index(&_gvn, idx, sizetype, ctrl);
1719 Node* scale = _gvn.transform( new LShiftXNode(idx, intcon(shift)) );
1720 return basic_plus_adr(ary, base, scale);
1721 }
1722
1723 //-------------------------load_array_element-------------------------
1724 Node* GraphKit::load_array_element(Node* ctl, Node* ary, Node* idx, const TypeAryPtr* arytype) {
1725 const Type* elemtype = arytype->elem();
1726 BasicType elembt = elemtype->array_element_basic_type();
1727 Node* adr = array_element_address(ary, idx, elembt, arytype->size());
1728 if (elembt == T_NARROWOOP) {
1729 elembt = T_OBJECT; // To satisfy switch in LoadNode::make()
1730 }
1731 Node* ld = make_load(ctl, adr, elemtype, elembt, arytype, MemNode::unordered);
1732 return ld;
1733 }
1734
1735 //-------------------------set_arguments_for_java_call-------------------------
1736 // Arguments (pre-popped from the stack) are taken from the JVMS.
1737 void GraphKit::set_arguments_for_java_call(CallJavaNode* call) {
1738 // Add the call arguments:
1739 uint nargs = call->method()->arg_size();
1740 for (uint i = 0; i < nargs; i++) {
1741 Node* arg = argument(i);
1742 call->init_req(i + TypeFunc::Parms, arg);
1743 }
1744 }
1745
1746 //---------------------------set_edges_for_java_call---------------------------
1747 // Connect a newly created call into the current JVMS.
1748 // A return value node (if any) is returned from set_edges_for_java_call.
1749 void GraphKit::set_edges_for_java_call(CallJavaNode* call, bool must_throw, bool separate_io_proj) {
1750
1751 // Add the predefined inputs:
1752 call->init_req( TypeFunc::Control, control() );
1753 call->init_req( TypeFunc::I_O , i_o() );
1754 call->init_req( TypeFunc::Memory , reset_memory() );
1755 call->init_req( TypeFunc::FramePtr, frameptr() );
1756 call->init_req( TypeFunc::ReturnAdr, top() );
1757
1758 add_safepoint_edges(call, must_throw);
1759
1760 Node* xcall = _gvn.transform(call);
1761
1762 if (xcall == top()) {
1763 set_control(top());
1764 return;
1765 }
1766 assert(xcall == call, "call identity is stable");
1767
1768 // Re-use the current map to produce the result.
1769
1770 set_control(_gvn.transform(new ProjNode(call, TypeFunc::Control)));
1771 set_i_o( _gvn.transform(new ProjNode(call, TypeFunc::I_O , separate_io_proj)));
1772 set_all_memory_call(xcall, separate_io_proj);
1773
1774 //return xcall; // no need, caller already has it
1775 }
1776
1777 Node* GraphKit::set_results_for_java_call(CallJavaNode* call, bool separate_io_proj, bool deoptimize) {
1778 if (stopped()) return top(); // maybe the call folded up?
1779
1780 // Capture the return value, if any.
1781 Node* ret;
1782 if (call->method() == NULL ||
1783 call->method()->return_type()->basic_type() == T_VOID)
1784 ret = top();
1785 else ret = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
1786
1787 // Note: Since any out-of-line call can produce an exception,
1788 // we always insert an I_O projection from the call into the result.
1789
1790 make_slow_call_ex(call, env()->Throwable_klass(), separate_io_proj, deoptimize);
1791
1792 if (separate_io_proj) {
1793 // The caller requested separate projections be used by the fall
1794 // through and exceptional paths, so replace the projections for
1795 // the fall through path.
1796 set_i_o(_gvn.transform( new ProjNode(call, TypeFunc::I_O) ));
1797 set_all_memory(_gvn.transform( new ProjNode(call, TypeFunc::Memory) ));
1798 }
1799 return ret;
1800 }
1801
1802 //--------------------set_predefined_input_for_runtime_call--------------------
1803 // Reading and setting the memory state is way conservative here.
1804 // The real problem is that I am not doing real Type analysis on memory,
1805 // so I cannot distinguish card mark stores from other stores. Across a GC
1806 // point the Store Barrier and the card mark memory has to agree. I cannot
1807 // have a card mark store and its barrier split across the GC point from
1808 // either above or below. Here I get that to happen by reading ALL of memory.
1809 // A better answer would be to separate out card marks from other memory.
1810 // For now, return the input memory state, so that it can be reused
1811 // after the call, if this call has restricted memory effects.
1812 Node* GraphKit::set_predefined_input_for_runtime_call(SafePointNode* call, Node* narrow_mem) {
1813 // Set fixed predefined input arguments
1814 Node* memory = reset_memory();
1815 Node* m = narrow_mem == NULL ? memory : narrow_mem;
1816 call->init_req( TypeFunc::Control, control() );
1817 call->init_req( TypeFunc::I_O, top() ); // does no i/o
1818 call->init_req( TypeFunc::Memory, m ); // may gc ptrs
1869 if (use->is_MergeMem()) {
1870 wl.push(use);
1871 }
1872 }
1873 }
1874
1875 // Replace the call with the current state of the kit.
1876 void GraphKit::replace_call(CallNode* call, Node* result, bool do_replaced_nodes) {
1877 JVMState* ejvms = NULL;
1878 if (has_exceptions()) {
1879 ejvms = transfer_exceptions_into_jvms();
1880 }
1881
1882 ReplacedNodes replaced_nodes = map()->replaced_nodes();
1883 ReplacedNodes replaced_nodes_exception;
1884 Node* ex_ctl = top();
1885
1886 SafePointNode* final_state = stop();
1887
1888 // Find all the needed outputs of this call
1889 CallProjections callprojs;
1890 call->extract_projections(&callprojs, true);
1891
1892 Unique_Node_List wl;
1893 Node* init_mem = call->in(TypeFunc::Memory);
1894 Node* final_mem = final_state->in(TypeFunc::Memory);
1895 Node* final_ctl = final_state->in(TypeFunc::Control);
1896 Node* final_io = final_state->in(TypeFunc::I_O);
1897
1898 // Replace all the old call edges with the edges from the inlining result
1899 if (callprojs.fallthrough_catchproj != NULL) {
1900 C->gvn_replace_by(callprojs.fallthrough_catchproj, final_ctl);
1901 }
1902 if (callprojs.fallthrough_memproj != NULL) {
1903 if (final_mem->is_MergeMem()) {
1904 // Parser's exits MergeMem was not transformed but may be optimized
1905 final_mem = _gvn.transform(final_mem);
1906 }
1907 C->gvn_replace_by(callprojs.fallthrough_memproj, final_mem);
1908 add_mergemem_users_to_worklist(wl, final_mem);
1909 }
1910 if (callprojs.fallthrough_ioproj != NULL) {
1911 C->gvn_replace_by(callprojs.fallthrough_ioproj, final_io);
1912 }
1913
1914 // Replace the result with the new result if it exists and is used
1915 if (callprojs.resproj != NULL && result != NULL) {
1916 C->gvn_replace_by(callprojs.resproj, result);
1917 }
1918
1919 if (ejvms == NULL) {
1920 // No exception edges to simply kill off those paths
1921 if (callprojs.catchall_catchproj != NULL) {
1922 C->gvn_replace_by(callprojs.catchall_catchproj, C->top());
1923 }
1924 if (callprojs.catchall_memproj != NULL) {
1925 C->gvn_replace_by(callprojs.catchall_memproj, C->top());
1926 }
1927 if (callprojs.catchall_ioproj != NULL) {
1928 C->gvn_replace_by(callprojs.catchall_ioproj, C->top());
1929 }
1930 // Replace the old exception object with top
1931 if (callprojs.exobj != NULL) {
1932 C->gvn_replace_by(callprojs.exobj, C->top());
1933 }
1934 } else {
1935 GraphKit ekit(ejvms);
1936
1937 // Load my combined exception state into the kit, with all phis transformed:
1938 SafePointNode* ex_map = ekit.combine_and_pop_all_exception_states();
1939 replaced_nodes_exception = ex_map->replaced_nodes();
1940
1941 Node* ex_oop = ekit.use_exception_state(ex_map);
1942
1943 if (callprojs.catchall_catchproj != NULL) {
1944 C->gvn_replace_by(callprojs.catchall_catchproj, ekit.control());
1945 ex_ctl = ekit.control();
1946 }
1947 if (callprojs.catchall_memproj != NULL) {
1948 Node* ex_mem = ekit.reset_memory();
1949 C->gvn_replace_by(callprojs.catchall_memproj, ex_mem);
1950 add_mergemem_users_to_worklist(wl, ex_mem);
1951 }
1952 if (callprojs.catchall_ioproj != NULL) {
1953 C->gvn_replace_by(callprojs.catchall_ioproj, ekit.i_o());
1954 }
1955
1956 // Replace the old exception object with the newly created one
1957 if (callprojs.exobj != NULL) {
1958 C->gvn_replace_by(callprojs.exobj, ex_oop);
1959 }
1960 }
1961
1962 // Disconnect the call from the graph
1963 call->disconnect_inputs(NULL, C);
1964 C->gvn_replace_by(call, C->top());
1965
1966 // Clean up any MergeMems that feed other MergeMems since the
1967 // optimizer doesn't like that.
1968 while (wl.size() > 0) {
1969 _gvn.transform(wl.pop());
1970 }
1971
1972 if (callprojs.fallthrough_catchproj != NULL && !final_ctl->is_top() && do_replaced_nodes) {
1973 replaced_nodes.apply(C, final_ctl);
1974 }
1975 if (!ex_ctl->is_top() && do_replaced_nodes) {
1976 replaced_nodes_exception.apply(C, ex_ctl);
1977 }
1978 }
1979
1980
1981 //------------------------------increment_counter------------------------------
1982 // for statistics: increment a VM counter by 1
1983
1984 void GraphKit::increment_counter(address counter_addr) {
1985 Node* adr1 = makecon(TypeRawPtr::make(counter_addr));
1986 increment_counter(adr1);
1987 }
1988
1989 void GraphKit::increment_counter(Node* counter_addr) {
1990 int adr_type = Compile::AliasIdxRaw;
1991 Node* ctrl = control();
1992 Node* cnt = make_load(ctrl, counter_addr, TypeInt::INT, T_INT, adr_type, MemNode::unordered);
2129 // it does not require card marks.
2130 Node* GraphKit::just_allocated_object(Node* current_control) {
2131 Node* ctrl = current_control;
2132 // Object::<init> is invoked after allocation, most of invoke nodes
2133 // will be reduced, but a region node is kept in parse time, we check
2134 // the pattern and skip the region node if it degraded to a copy.
2135 if (ctrl != NULL && ctrl->is_Region() && ctrl->req() == 2 &&
2136 ctrl->as_Region()->is_copy()) {
2137 ctrl = ctrl->as_Region()->is_copy();
2138 }
2139 if (C->recent_alloc_ctl() == ctrl) {
2140 return C->recent_alloc_obj();
2141 }
2142 return NULL;
2143 }
2144
2145
2146 void GraphKit::round_double_arguments(ciMethod* dest_method) {
2147 // (Note: TypeFunc::make has a cache that makes this fast.)
2148 const TypeFunc* tf = TypeFunc::make(dest_method);
2149 int nargs = tf->domain()->cnt() - TypeFunc::Parms;
2150 for (int j = 0; j < nargs; j++) {
2151 const Type *targ = tf->domain()->field_at(j + TypeFunc::Parms);
2152 if( targ->basic_type() == T_DOUBLE ) {
2153 // If any parameters are doubles, they must be rounded before
2154 // the call, dstore_rounding does gvn.transform
2155 Node *arg = argument(j);
2156 arg = dstore_rounding(arg);
2157 set_argument(j, arg);
2158 }
2159 }
2160 }
2161
2162 /**
2163 * Record profiling data exact_kls for Node n with the type system so
2164 * that it can propagate it (speculation)
2165 *
2166 * @param n node that the type applies to
2167 * @param exact_kls type from profiling
2168 * @param maybe_null did profiling see null?
2169 *
2170 * @return node with improved type
2171 */
2188 speculative = speculative->with_inline_depth(jvms()->depth());
2189 } else if (current_type->would_improve_ptr(ptr_kind)) {
2190 // Profiling report that null was never seen so we can change the
2191 // speculative type to non null ptr.
2192 if (ptr_kind == ProfileAlwaysNull) {
2193 speculative = TypePtr::NULL_PTR;
2194 } else {
2195 assert(ptr_kind == ProfileNeverNull, "nothing else is an improvement");
2196 const TypePtr* ptr = TypePtr::NOTNULL;
2197 if (speculative != NULL) {
2198 speculative = speculative->cast_to_ptr_type(ptr->ptr())->is_ptr();
2199 } else {
2200 speculative = ptr;
2201 }
2202 }
2203 }
2204
2205 if (speculative != current_type->speculative()) {
2206 // Build a type with a speculative type (what we think we know
2207 // about the type but will need a guard when we use it)
2208 const TypeOopPtr* spec_type = TypeOopPtr::make(TypePtr::BotPTR, Type::OffsetBot, TypeOopPtr::InstanceBot, speculative);
2209 // We're changing the type, we need a new CheckCast node to carry
2210 // the new type. The new type depends on the control: what
2211 // profiling tells us is only valid from here as far as we can
2212 // tell.
2213 Node* cast = new CheckCastPPNode(control(), n, current_type->remove_speculative()->join_speculative(spec_type));
2214 cast = _gvn.transform(cast);
2215 replace_in_map(n, cast);
2216 n = cast;
2217 }
2218
2219 return n;
2220 }
2221
2222 /**
2223 * Record profiling data from receiver profiling at an invoke with the
2224 * type system so that it can propagate it (speculation)
2225 *
2226 * @param n receiver node
2227 *
2228 * @return node with improved type
2253 }
2254 ptr_kind = (i == call->row_limit()) ? ProfileAlwaysNull : ProfileMaybeNull;
2255 }
2256 }
2257 }
2258 return record_profile_for_speculation(n, exact_kls, ptr_kind);
2259 }
2260
2261 /**
2262 * Record profiling data from argument profiling at an invoke with the
2263 * type system so that it can propagate it (speculation)
2264 *
2265 * @param dest_method target method for the call
2266 * @param bc what invoke bytecode is this?
2267 */
2268 void GraphKit::record_profiled_arguments_for_speculation(ciMethod* dest_method, Bytecodes::Code bc) {
2269 if (!UseTypeSpeculation) {
2270 return;
2271 }
2272 const TypeFunc* tf = TypeFunc::make(dest_method);
2273 int nargs = tf->domain()->cnt() - TypeFunc::Parms;
2274 int skip = Bytecodes::has_receiver(bc) ? 1 : 0;
2275 for (int j = skip, i = 0; j < nargs && i < TypeProfileArgsLimit; j++) {
2276 const Type *targ = tf->domain()->field_at(j + TypeFunc::Parms);
2277 if (targ->basic_type() == T_OBJECT || targ->basic_type() == T_ARRAY) {
2278 ProfilePtrKind ptr_kind = ProfileMaybeNull;
2279 ciKlass* better_type = NULL;
2280 if (method()->argument_profiled_type(bci(), i, better_type, ptr_kind)) {
2281 record_profile_for_speculation(argument(j), better_type, ptr_kind);
2282 }
2283 i++;
2284 }
2285 }
2286 }
2287
2288 /**
2289 * Record profiling data from parameter profiling at an invoke with
2290 * the type system so that it can propagate it (speculation)
2291 */
2292 void GraphKit::record_profiled_parameters_for_speculation() {
2293 if (!UseTypeSpeculation) {
2294 return;
2295 }
2296 for (int i = 0, j = 0; i < method()->arg_size() ; i++) {
2297 if (_gvn.type(local(i))->isa_oopptr()) {
2772 // The decision to inline or out-of-line this final check is platform
2773 // dependent, and is found in the AD file definition of PartialSubtypeCheck.
2774 Node* psc = gvn->transform(
2775 new PartialSubtypeCheckNode(*ctrl, subklass, superklass));
2776
2777 IfNode *iff4 = gen_subtype_check_compare(*ctrl, psc, gvn->zerocon(T_OBJECT), BoolTest::ne, PROB_FAIR, gvn, T_ADDRESS);
2778 r_not_subtype->init_req(2, gvn->transform(new IfTrueNode (iff4)));
2779 r_ok_subtype ->init_req(3, gvn->transform(new IfFalseNode(iff4)));
2780
2781 // Return false path; set default control to true path.
2782 *ctrl = gvn->transform(r_ok_subtype);
2783 return gvn->transform(r_not_subtype);
2784 }
2785
2786 // Profile-driven exact type check:
2787 Node* GraphKit::type_check_receiver(Node* receiver, ciKlass* klass,
2788 float prob,
2789 Node* *casted_receiver) {
2790 const TypeKlassPtr* tklass = TypeKlassPtr::make(klass);
2791 Node* recv_klass = load_object_klass(receiver);
2792 Node* want_klass = makecon(tklass);
2793 Node* cmp = _gvn.transform( new CmpPNode(recv_klass, want_klass) );
2794 Node* bol = _gvn.transform( new BoolNode(cmp, BoolTest::eq) );
2795 IfNode* iff = create_and_xform_if(control(), bol, prob, COUNT_UNKNOWN);
2796 set_control( _gvn.transform( new IfTrueNode (iff) ));
2797 Node* fail = _gvn.transform( new IfFalseNode(iff) );
2798
2799 const TypeOopPtr* recv_xtype = tklass->as_instance_type();
2800 assert(recv_xtype->klass_is_exact(), "");
2801
2802 // Subsume downstream occurrences of receiver with a cast to
2803 // recv_xtype, since now we know what the type will be.
2804 Node* cast = new CheckCastPPNode(control(), receiver, recv_xtype);
2805 (*casted_receiver) = _gvn.transform(cast);
2806 // (User must make the replace_in_map call.)
2807
2808 return fail;
2809 }
2810
2811 //------------------------------subtype_check_receiver-------------------------
2812 Node* GraphKit::subtype_check_receiver(Node* receiver, ciKlass* klass,
2813 Node** casted_receiver) {
2814 const TypeKlassPtr* tklass = TypeKlassPtr::make(klass);
2815 Node* recv_klass = load_object_klass(receiver);
2816 Node* want_klass = makecon(tklass);
2817
2818 Node* slow_ctl = gen_subtype_check(recv_klass, want_klass);
2819
2820 // Cast receiver after successful check
2821 const TypeOopPtr* recv_type = tklass->cast_to_exactness(false)->is_klassptr()->as_instance_type();
2822 Node* cast = new CheckCastPPNode(control(), receiver, recv_type);
2823 (*casted_receiver) = _gvn.transform(cast);
2824
2825 return slow_ctl;
2826 }
2827
2828 //------------------------------seems_never_null-------------------------------
2829 // Use null_seen information if it is available from the profile.
2830 // If we see an unexpected null at a type check we record it and force a
3015 // and the reflective instance-of call.
3016 Node* GraphKit::gen_instanceof(Node* obj, Node* superklass, bool safe_for_replace) {
3017 kill_dead_locals(); // Benefit all the uncommon traps
3018 assert( !stopped(), "dead parse path should be checked in callers" );
3019 assert(!TypePtr::NULL_PTR->higher_equal(_gvn.type(superklass)->is_klassptr()),
3020 "must check for not-null not-dead klass in callers");
3021
3022 // Make the merge point
3023 enum { _obj_path = 1, _fail_path, _null_path, PATH_LIMIT };
3024 RegionNode* region = new RegionNode(PATH_LIMIT);
3025 Node* phi = new PhiNode(region, TypeInt::BOOL);
3026 C->set_has_split_ifs(true); // Has chance for split-if optimization
3027
3028 ciProfileData* data = NULL;
3029 if (java_bc() == Bytecodes::_instanceof) { // Only for the bytecode
3030 data = method()->method_data()->bci_to_data(bci());
3031 }
3032 bool speculative_not_null = false;
3033 bool never_see_null = (ProfileDynamicTypes // aggressive use of profile
3034 && seems_never_null(obj, data, speculative_not_null));
3035
3036 // Null check; get casted pointer; set region slot 3
3037 Node* null_ctl = top();
3038 Node* not_null_obj = null_check_oop(obj, &null_ctl, never_see_null, safe_for_replace, speculative_not_null);
3039
3040 // If not_null_obj is dead, only null-path is taken
3041 if (stopped()) { // Doing instance-of on a NULL?
3042 set_control(null_ctl);
3043 return intcon(0);
3044 }
3045 region->init_req(_null_path, null_ctl);
3046 phi ->init_req(_null_path, intcon(0)); // Set null path value
3047 if (null_ctl == top()) {
3048 // Do this eagerly, so that pattern matches like is_diamond_phi
3049 // will work even during parsing.
3050 assert(_null_path == PATH_LIMIT-1, "delete last");
3051 region->del_req(_null_path);
3052 phi ->del_req(_null_path);
3053 }
3054
3055 // Do we know the type check always succeed?
3056 bool known_statically = false;
3057 if (_gvn.type(superklass)->singleton()) {
3058 ciKlass* superk = _gvn.type(superklass)->is_klassptr()->klass();
3059 ciKlass* subk = _gvn.type(obj)->is_oopptr()->klass();
3060 if (subk != NULL && subk->is_loaded()) {
3061 int static_res = C->static_subtype_check(superk, subk);
3062 known_statically = (static_res == Compile::SSC_always_true || static_res == Compile::SSC_always_false);
3063 }
3064 }
3065
3066 if (!known_statically) {
3067 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr();
3068 // We may not have profiling here or it may not help us. If we
3069 // have a speculative type use it to perform an exact cast.
3070 ciKlass* spec_obj_type = obj_type->speculative_type();
3071 if (spec_obj_type != NULL || (ProfileDynamicTypes && data != NULL)) {
3072 Node* cast_obj = maybe_cast_profiled_receiver(not_null_obj, NULL, spec_obj_type, safe_for_replace);
3073 if (stopped()) { // Profile disagrees with this path.
3074 set_control(null_ctl); // Null is the only remaining possibility.
3075 return intcon(0);
3076 }
3077 if (cast_obj != NULL) {
3078 not_null_obj = cast_obj;
3079 }
3080 }
3081 }
3082
3083 // Load the object's klass
3084 Node* obj_klass = load_object_klass(not_null_obj);
3085
3086 // Generate the subtype check
3087 Node* not_subtype_ctrl = gen_subtype_check(obj_klass, superklass);
3088
3089 // Plug in the success path to the general merge in slot 1.
3090 region->init_req(_obj_path, control());
3091 phi ->init_req(_obj_path, intcon(1));
3092
3093 // Plug in the failing path to the general merge in slot 2.
3094 region->init_req(_fail_path, not_subtype_ctrl);
3095 phi ->init_req(_fail_path, intcon(0));
3096
3097 // Return final merged results
3098 set_control( _gvn.transform(region) );
3099 record_for_igvn(region);
3100
3101 // If we know the type check always succeeds then we don't use the
3102 // profiling data at this bytecode. Don't lose it, feed it to the
3103 // type system as a speculative type.
3104 if (safe_for_replace) {
3105 Node* casted_obj = record_profiled_receiver_for_speculation(obj);
3106 replace_in_map(obj, casted_obj);
3107 }
3108
3109 return _gvn.transform(phi);
3110 }
3111
3112 //-------------------------------gen_checkcast---------------------------------
3113 // Generate a checkcast idiom. Used by both the checkcast bytecode and the
3114 // array store bytecode. Stack must be as-if BEFORE doing the bytecode so the
3115 // uncommon-trap paths work. Adjust stack after this call.
3116 // If failure_control is supplied and not null, it is filled in with
3117 // the control edge for the cast failure. Otherwise, an appropriate
3118 // uncommon trap or exception is thrown.
3119 Node* GraphKit::gen_checkcast(Node *obj, Node* superklass,
3120 Node* *failure_control) {
3121 kill_dead_locals(); // Benefit all the uncommon traps
3122 const TypeKlassPtr *tk = _gvn.type(superklass)->is_klassptr();
3123 const Type *toop = TypeOopPtr::make_from_klass(tk->klass());
3124
3125 // Fast cutout: Check the case that the cast is vacuously true.
3126 // This detects the common cases where the test will short-circuit
3127 // away completely. We do this before we perform the null check,
3128 // because if the test is going to turn into zero code, we don't
3129 // want a residual null check left around. (Causes a slowdown,
3130 // for example, in some objArray manipulations, such as a[i]=a[j].)
3131 if (tk->singleton()) {
3132 const TypeOopPtr* objtp = _gvn.type(obj)->isa_oopptr();
3133 if (objtp != NULL && objtp->klass() != NULL) {
3134 switch (C->static_subtype_check(tk->klass(), objtp->klass())) {
3135 case Compile::SSC_always_true:
3136 // If we know the type check always succeed then we don't use
3137 // the profiling data at this bytecode. Don't lose it, feed it
3138 // to the type system as a speculative type.
3139 return record_profiled_receiver_for_speculation(obj);
3140 case Compile::SSC_always_false:
3141 // It needs a null check because a null will *pass* the cast check.
3142 // A non-null value will always produce an exception.
3143 return null_assert(obj);
3144 }
3145 }
3146 }
3147
3148 ciProfileData* data = NULL;
3149 bool safe_for_replace = false;
3150 if (failure_control == NULL) { // use MDO in regular case only
3151 assert(java_bc() == Bytecodes::_aastore ||
3152 java_bc() == Bytecodes::_checkcast,
3153 "interpreter profiles type checks only for these BCs");
3154 data = method()->method_data()->bci_to_data(bci());
3155 safe_for_replace = true;
3156 }
3157
3158 // Make the merge point
3159 enum { _obj_path = 1, _null_path, PATH_LIMIT };
3160 RegionNode* region = new RegionNode(PATH_LIMIT);
3161 Node* phi = new PhiNode(region, toop);
3162 C->set_has_split_ifs(true); // Has chance for split-if optimization
3163
3164 // Use null-cast information if it is available
3165 bool speculative_not_null = false;
3166 bool never_see_null = ((failure_control == NULL) // regular case only
3167 && seems_never_null(obj, data, speculative_not_null));
3168
3169 // Null check; get casted pointer; set region slot 3
3170 Node* null_ctl = top();
3171 Node* not_null_obj = null_check_oop(obj, &null_ctl, never_see_null, safe_for_replace, speculative_not_null);
3172
3173 // If not_null_obj is dead, only null-path is taken
3174 if (stopped()) { // Doing instance-of on a NULL?
3175 set_control(null_ctl);
3176 return null();
3177 }
3178 region->init_req(_null_path, null_ctl);
3179 phi ->init_req(_null_path, null()); // Set null path value
3180 if (null_ctl == top()) {
3181 // Do this eagerly, so that pattern matches like is_diamond_phi
3182 // will work even during parsing.
3183 assert(_null_path == PATH_LIMIT-1, "delete last");
3184 region->del_req(_null_path);
3185 phi ->del_req(_null_path);
3186 }
3187
3188 Node* cast_obj = NULL;
3189 if (tk->klass_is_exact()) {
3190 // The following optimization tries to statically cast the speculative type of the object
3191 // (for example obtained during profiling) to the type of the superklass and then do a
3192 // dynamic check that the type of the object is what we expect. To work correctly
3193 // for checkcast and aastore the type of superklass should be exact.
3194 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr();
3195 // We may not have profiling here or it may not help us. If we have
3196 // a speculative type use it to perform an exact cast.
3197 ciKlass* spec_obj_type = obj_type->speculative_type();
3198 if (spec_obj_type != NULL || data != NULL) {
3199 cast_obj = maybe_cast_profiled_receiver(not_null_obj, tk->klass(), spec_obj_type, safe_for_replace);
3200 if (cast_obj != NULL) {
3201 if (failure_control != NULL) // failure is now impossible
3202 (*failure_control) = top();
3203 // adjust the type of the phi to the exact klass:
3204 phi->raise_bottom_type(_gvn.type(cast_obj)->meet_speculative(TypePtr::NULL_PTR));
3205 }
3206 }
3207 }
3208
3209 if (cast_obj == NULL) {
3210 // Load the object's klass
3211 Node* obj_klass = load_object_klass(not_null_obj);
3212
3213 // Generate the subtype check
3214 Node* not_subtype_ctrl = gen_subtype_check( obj_klass, superklass );
3215
3216 // Plug in success path into the merge
3217 cast_obj = _gvn.transform(new CheckCastPPNode(control(), not_null_obj, toop));
3218 // Failure path ends in uncommon trap (or may be dead - failure impossible)
3219 if (failure_control == NULL) {
3220 if (not_subtype_ctrl != top()) { // If failure is possible
3221 PreserveJVMState pjvms(this);
3222 set_control(not_subtype_ctrl);
3223 builtin_throw(Deoptimization::Reason_class_check, obj_klass);
3224 }
3225 } else {
3226 (*failure_control) = not_subtype_ctrl;
3227 }
3228 }
3229
3230 region->init_req(_obj_path, control());
3231 phi ->init_req(_obj_path, cast_obj);
3232
3233 // A merge of NULL or Casted-NotNull obj
3234 Node* res = _gvn.transform(phi);
3235
3236 // Note I do NOT always 'replace_in_map(obj,result)' here.
3237 // if( tk->klass()->can_be_primary_super() )
3238 // This means that if I successfully store an Object into an array-of-String
3239 // I 'forget' that the Object is really now known to be a String. I have to
3240 // do this because we don't have true union types for interfaces - if I store
3241 // a Baz into an array-of-Interface and then tell the optimizer it's an
3242 // Interface, I forget that it's also a Baz and cannot do Baz-like field
3243 // references to it. FIX THIS WHEN UNION TYPES APPEAR!
3244 // replace_in_map( obj, res );
3245
3246 // Return final merged results
3247 set_control( _gvn.transform(region) );
3248 record_for_igvn(region);
3249
3250 return record_profiled_receiver_for_speculation(res);
3251 }
3252
3253 //------------------------------next_monitor-----------------------------------
3254 // What number should be given to the next monitor?
3255 int GraphKit::next_monitor() {
3256 int current = jvms()->monitor_depth()* C->sync_stack_slots();
3257 int next = current + C->sync_stack_slots();
3258 // Keep the toplevel high water mark current:
3259 if (C->fixed_slots() < next) C->set_fixed_slots(next);
3260 return current;
3261 }
3262
3263 //------------------------------insert_mem_bar---------------------------------
3264 // Memory barrier to avoid floating things around
3265 // The membar serves as a pinch point between both control and all memory slices.
3266 Node* GraphKit::insert_mem_bar(int opcode, Node* precedent) {
3267 MemBarNode* mb = MemBarNode::make(C, opcode, Compile::AliasIdxBot, precedent);
3268 mb->init_req(TypeFunc::Control, control());
3269 mb->init_req(TypeFunc::Memory, reset_memory());
3270 Node* membar = _gvn.transform(mb);
3271 set_control(_gvn.transform(new ProjNode(membar, TypeFunc::Control)));
3272 set_all_memory_call(membar);
3310 Node* mem = reset_memory();
3311 MemBarNode* mb = MemBarNode::make(C, Op_MemBarVolatile, Compile::AliasIdxRaw);
3312 mb->init_req(TypeFunc::Control, control());
3313 mb->init_req(TypeFunc::Memory, mem);
3314 Node* membar = _gvn.transform(mb);
3315 set_control(_gvn.transform(new ProjNode(membar, TypeFunc::Control)));
3316 Node* newmem = _gvn.transform(new ProjNode(membar, TypeFunc::Memory));
3317 set_all_memory(mem);
3318 set_memory(newmem, Compile::AliasIdxRaw);
3319 }
3320
3321 //------------------------------shared_lock------------------------------------
3322 // Emit locking code.
3323 FastLockNode* GraphKit::shared_lock(Node* obj) {
3324 // bci is either a monitorenter bc or InvocationEntryBci
3325 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces
3326 assert(SynchronizationEntryBCI == InvocationEntryBci, "");
3327
3328 if( !GenerateSynchronizationCode )
3329 return NULL; // Not locking things?
3330 if (stopped()) // Dead monitor?
3331 return NULL;
3332
3333 assert(dead_locals_are_killed(), "should kill locals before sync. point");
3334
3335 obj = access_resolve(obj, ACCESS_READ | ACCESS_WRITE);
3336
3337 // Box the stack location
3338 Node* box = _gvn.transform(new BoxLockNode(next_monitor()));
3339 Node* mem = reset_memory();
3340
3341 FastLockNode * flock = _gvn.transform(new FastLockNode(0, obj, box) )->as_FastLock();
3342 if (UseBiasedLocking && PrintPreciseBiasedLockingStatistics) {
3343 // Create the counters for this fast lock.
3344 flock->create_lock_counter(sync_jvms()); // sync_jvms used to get current bci
3345 }
3346
3347 // Create the rtm counters for this fast lock if needed.
3348 flock->create_rtm_lock_counter(sync_jvms()); // sync_jvms used to get current bci
3349
3384 }
3385 #endif
3386
3387 return flock;
3388 }
3389
3390
3391 //------------------------------shared_unlock----------------------------------
3392 // Emit unlocking code.
3393 void GraphKit::shared_unlock(Node* box, Node* obj) {
3394 // bci is either a monitorenter bc or InvocationEntryBci
3395 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces
3396 assert(SynchronizationEntryBCI == InvocationEntryBci, "");
3397
3398 if( !GenerateSynchronizationCode )
3399 return;
3400 if (stopped()) { // Dead monitor?
3401 map()->pop_monitor(); // Kill monitor from debug info
3402 return;
3403 }
3404
3405 // Memory barrier to avoid floating things down past the locked region
3406 insert_mem_bar(Op_MemBarReleaseLock);
3407
3408 const TypeFunc *tf = OptoRuntime::complete_monitor_exit_Type();
3409 UnlockNode *unlock = new UnlockNode(C, tf);
3410 #ifdef ASSERT
3411 unlock->set_dbg_jvms(sync_jvms());
3412 #endif
3413 uint raw_idx = Compile::AliasIdxRaw;
3414 unlock->init_req( TypeFunc::Control, control() );
3415 unlock->init_req( TypeFunc::Memory , memory(raw_idx) );
3416 unlock->init_req( TypeFunc::I_O , top() ) ; // does no i/o
3417 unlock->init_req( TypeFunc::FramePtr, frameptr() );
3418 unlock->init_req( TypeFunc::ReturnAdr, top() );
3419
3420 unlock->init_req(TypeFunc::Parms + 0, obj);
3421 unlock->init_req(TypeFunc::Parms + 1, box);
3422 unlock = _gvn.transform(unlock)->as_Unlock();
3423
3424 Node* mem = reset_memory();
3425
3426 // unlock has no side-effects, sets few values
3427 set_predefined_output_for_runtime_call(unlock, mem, TypeRawPtr::BOTTOM);
3428
3429 // Kill monitor from debug info
3430 map()->pop_monitor( );
3431 }
3432
3433 //-------------------------------get_layout_helper-----------------------------
3434 // If the given klass is a constant or known to be an array,
3435 // fetch the constant layout helper value into constant_value
3436 // and return (Node*)NULL. Otherwise, load the non-constant
3437 // layout helper value, and return the node which represents it.
3438 // This two-faced routine is useful because allocation sites
3439 // almost always feature constant types.
3440 Node* GraphKit::get_layout_helper(Node* klass_node, jint& constant_value) {
3441 const TypeKlassPtr* inst_klass = _gvn.type(klass_node)->isa_klassptr();
3442 if (!StressReflectiveCode && inst_klass != NULL) {
3443 ciKlass* klass = inst_klass->klass();
3444 bool xklass = inst_klass->klass_is_exact();
3445 if (xklass || klass->is_array_klass()) {
3446 jint lhelper = klass->layout_helper();
3447 if (lhelper != Klass::_lh_neutral_value) {
3448 constant_value = lhelper;
3449 return (Node*) NULL;
3450 }
3451 }
3452 }
3453 constant_value = Klass::_lh_neutral_value; // put in a known value
3454 Node* lhp = basic_plus_adr(klass_node, klass_node, in_bytes(Klass::layout_helper_offset()));
3455 return make_load(NULL, lhp, TypeInt::INT, T_INT, MemNode::unordered);
3456 }
3457
3458 // We just put in an allocate/initialize with a big raw-memory effect.
3459 // Hook selected additional alias categories on the initialization.
3460 static void hook_memory_on_init(GraphKit& kit, int alias_idx,
3461 MergeMemNode* init_in_merge,
3462 Node* init_out_raw) {
3463 DEBUG_ONLY(Node* init_in_raw = init_in_merge->base_memory());
3464 assert(init_in_merge->memory_at(alias_idx) == init_in_raw, "");
3465
3487
3488 // a normal slow-call doesn't change i_o, but an allocation does
3489 // we create a separate i_o projection for the normal control path
3490 set_i_o(_gvn.transform( new ProjNode(allocx, TypeFunc::I_O, false) ) );
3491 Node* rawoop = _gvn.transform( new ProjNode(allocx, TypeFunc::Parms) );
3492
3493 // put in an initialization barrier
3494 InitializeNode* init = insert_mem_bar_volatile(Op_Initialize, rawidx,
3495 rawoop)->as_Initialize();
3496 assert(alloc->initialization() == init, "2-way macro link must work");
3497 assert(init ->allocation() == alloc, "2-way macro link must work");
3498 {
3499 // Extract memory strands which may participate in the new object's
3500 // initialization, and source them from the new InitializeNode.
3501 // This will allow us to observe initializations when they occur,
3502 // and link them properly (as a group) to the InitializeNode.
3503 assert(init->in(InitializeNode::Memory) == malloc, "");
3504 MergeMemNode* minit_in = MergeMemNode::make(malloc);
3505 init->set_req(InitializeNode::Memory, minit_in);
3506 record_for_igvn(minit_in); // fold it up later, if possible
3507 Node* minit_out = memory(rawidx);
3508 assert(minit_out->is_Proj() && minit_out->in(0) == init, "");
3509 // Add an edge in the MergeMem for the header fields so an access
3510 // to one of those has correct memory state
3511 set_memory(minit_out, C->get_alias_index(oop_type->add_offset(oopDesc::mark_offset_in_bytes())));
3512 set_memory(minit_out, C->get_alias_index(oop_type->add_offset(oopDesc::klass_offset_in_bytes())));
3513 if (oop_type->isa_aryptr()) {
3514 const TypePtr* telemref = oop_type->add_offset(Type::OffsetBot);
3515 int elemidx = C->get_alias_index(telemref);
3516 hook_memory_on_init(*this, elemidx, minit_in, minit_out);
3517 } else if (oop_type->isa_instptr()) {
3518 ciInstanceKlass* ik = oop_type->klass()->as_instance_klass();
3519 for (int i = 0, len = ik->nof_nonstatic_fields(); i < len; i++) {
3520 ciField* field = ik->nonstatic_field_at(i);
3521 if (field->offset() >= TrackedInitializationLimit * HeapWordSize)
3522 continue; // do not bother to track really large numbers of fields
3523 // Find (or create) the alias category for this field:
3524 int fieldidx = C->alias_type(field)->index();
3525 hook_memory_on_init(*this, fieldidx, minit_in, minit_out);
3526 }
3527 }
3528 }
3529
3530 // Cast raw oop to the real thing...
3531 Node* javaoop = new CheckCastPPNode(control(), rawoop, oop_type);
3532 javaoop = _gvn.transform(javaoop);
3533 C->set_recent_alloc(control(), javaoop);
3534 assert(just_allocated_object(control()) == javaoop, "just allocated");
3535
3536 #ifdef ASSERT
3537 { // Verify that the AllocateNode::Ideal_allocation recognizers work:
3548 assert(alloc->in(AllocateNode::ALength)->is_top(), "no length, please");
3549 }
3550 }
3551 #endif //ASSERT
3552
3553 return javaoop;
3554 }
3555
3556 //---------------------------new_instance--------------------------------------
3557 // This routine takes a klass_node which may be constant (for a static type)
3558 // or may be non-constant (for reflective code). It will work equally well
3559 // for either, and the graph will fold nicely if the optimizer later reduces
3560 // the type to a constant.
3561 // The optional arguments are for specialized use by intrinsics:
3562 // - If 'extra_slow_test' if not null is an extra condition for the slow-path.
3563 // - If 'return_size_val', report the the total object size to the caller.
3564 // - deoptimize_on_exception controls how Java exceptions are handled (rethrow vs deoptimize)
3565 Node* GraphKit::new_instance(Node* klass_node,
3566 Node* extra_slow_test,
3567 Node* *return_size_val,
3568 bool deoptimize_on_exception) {
3569 // Compute size in doublewords
3570 // The size is always an integral number of doublewords, represented
3571 // as a positive bytewise size stored in the klass's layout_helper.
3572 // The layout_helper also encodes (in a low bit) the need for a slow path.
3573 jint layout_con = Klass::_lh_neutral_value;
3574 Node* layout_val = get_layout_helper(klass_node, layout_con);
3575 int layout_is_con = (layout_val == NULL);
3576
3577 if (extra_slow_test == NULL) extra_slow_test = intcon(0);
3578 // Generate the initial go-slow test. It's either ALWAYS (return a
3579 // Node for 1) or NEVER (return a NULL) or perhaps (in the reflective
3580 // case) a computed value derived from the layout_helper.
3581 Node* initial_slow_test = NULL;
3582 if (layout_is_con) {
3583 assert(!StressReflectiveCode, "stress mode does not use these paths");
3584 bool must_go_slow = Klass::layout_helper_needs_slow_path(layout_con);
3585 initial_slow_test = must_go_slow ? intcon(1) : extra_slow_test;
3586 } else { // reflective case
3587 // This reflective path is used by Unsafe.allocateInstance.
3588 // (It may be stress-tested by specifying StressReflectiveCode.)
3589 // Basically, we want to get into the VM is there's an illegal argument.
3590 Node* bit = intcon(Klass::_lh_instance_slow_path_bit);
3591 initial_slow_test = _gvn.transform( new AndINode(layout_val, bit) );
3592 if (extra_slow_test != intcon(0)) {
3593 initial_slow_test = _gvn.transform( new OrINode(initial_slow_test, extra_slow_test) );
3594 }
3595 // (Macro-expander will further convert this to a Bool, if necessary.)
3606
3607 // Clear the low bits to extract layout_helper_size_in_bytes:
3608 assert((int)Klass::_lh_instance_slow_path_bit < BytesPerLong, "clear bit");
3609 Node* mask = MakeConX(~ (intptr_t)right_n_bits(LogBytesPerLong));
3610 size = _gvn.transform( new AndXNode(size, mask) );
3611 }
3612 if (return_size_val != NULL) {
3613 (*return_size_val) = size;
3614 }
3615
3616 // This is a precise notnull oop of the klass.
3617 // (Actually, it need not be precise if this is a reflective allocation.)
3618 // It's what we cast the result to.
3619 const TypeKlassPtr* tklass = _gvn.type(klass_node)->isa_klassptr();
3620 if (!tklass) tklass = TypeKlassPtr::OBJECT;
3621 const TypeOopPtr* oop_type = tklass->as_instance_type();
3622
3623 // Now generate allocation code
3624
3625 // The entire memory state is needed for slow path of the allocation
3626 // since GC and deoptimization can happened.
3627 Node *mem = reset_memory();
3628 set_all_memory(mem); // Create new memory state
3629
3630 AllocateNode* alloc = new AllocateNode(C, AllocateNode::alloc_type(Type::TOP),
3631 control(), mem, i_o(),
3632 size, klass_node,
3633 initial_slow_test);
3634
3635 return set_output_for_allocation(alloc, oop_type, deoptimize_on_exception);
3636 }
3637
3638 //-------------------------------new_array-------------------------------------
3639 // helper for both newarray and anewarray
3640 // The 'length' parameter is (obviously) the length of the array.
3641 // See comments on new_instance for the meaning of the other arguments.
3642 Node* GraphKit::new_array(Node* klass_node, // array klass (maybe variable)
3643 Node* length, // number of array elements
3644 int nargs, // number of arguments to push back for uncommon trap
3645 Node* *return_size_val,
3646 bool deoptimize_on_exception) {
3647 jint layout_con = Klass::_lh_neutral_value;
3648 Node* layout_val = get_layout_helper(klass_node, layout_con);
3649 int layout_is_con = (layout_val == NULL);
3650
3651 if (!layout_is_con && !StressReflectiveCode &&
3652 !too_many_traps(Deoptimization::Reason_class_check)) {
3653 // This is a reflective array creation site.
3654 // Optimistically assume that it is a subtype of Object[],
3655 // so that we can fold up all the address arithmetic.
3656 layout_con = Klass::array_layout_helper(T_OBJECT);
3657 Node* cmp_lh = _gvn.transform( new CmpINode(layout_val, intcon(layout_con)) );
3658 Node* bol_lh = _gvn.transform( new BoolNode(cmp_lh, BoolTest::eq) );
3659 { BuildCutout unless(this, bol_lh, PROB_MAX);
3660 inc_sp(nargs);
3661 uncommon_trap(Deoptimization::Reason_class_check,
3662 Deoptimization::Action_maybe_recompile);
3663 }
3664 layout_val = NULL;
3665 layout_is_con = true;
3666 }
3667
3668 // Generate the initial go-slow test. Make sure we do not overflow
3669 // if length is huge (near 2Gig) or negative! We do not need
3670 // exact double-words here, just a close approximation of needed
3671 // double-words. We can't add any offset or rounding bits, lest we
3672 // take a size -1 of bytes and make it positive. Use an unsigned
3673 // compare, so negative sizes look hugely positive.
3674 int fast_size_limit = FastAllocateSizeLimit;
3675 if (layout_is_con) {
3676 assert(!StressReflectiveCode, "stress mode does not use these paths");
3677 // Increase the size limit if we have exact knowledge of array type.
3678 int log2_esize = Klass::layout_helper_log2_element_size(layout_con);
3679 fast_size_limit <<= (LogBytesPerLong - log2_esize);
3680 }
3681
3682 Node* initial_slow_cmp = _gvn.transform( new CmpUNode( length, intcon( fast_size_limit ) ) );
3683 Node* initial_slow_test = _gvn.transform( new BoolNode( initial_slow_cmp, BoolTest::gt ) );
3684
3685 // --- Size Computation ---
3686 // array_size = round_to_heap(array_header + (length << elem_shift));
3687 // where round_to_heap(x) == align_to(x, MinObjAlignmentInBytes)
3688 // and align_to(x, y) == ((x + y-1) & ~(y-1))
3689 // The rounding mask is strength-reduced, if possible.
3690 int round_mask = MinObjAlignmentInBytes - 1;
3691 Node* header_size = NULL;
3692 int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE);
3693 // (T_BYTE has the weakest alignment and size restrictions...)
3694 if (layout_is_con) {
3695 int hsize = Klass::layout_helper_header_size(layout_con);
3696 int eshift = Klass::layout_helper_log2_element_size(layout_con);
3697 BasicType etype = Klass::layout_helper_element_type(layout_con);
3698 if ((round_mask & ~right_n_bits(eshift)) == 0)
3699 round_mask = 0; // strength-reduce it if it goes away completely
3700 assert((hsize & right_n_bits(eshift)) == 0, "hsize is pre-rounded");
3701 assert(header_size_min <= hsize, "generic minimum is smallest");
3702 header_size_min = hsize;
3703 header_size = intcon(hsize + round_mask);
3704 } else {
3705 Node* hss = intcon(Klass::_lh_header_size_shift);
3706 Node* hsm = intcon(Klass::_lh_header_size_mask);
3707 Node* hsize = _gvn.transform( new URShiftINode(layout_val, hss) );
3708 hsize = _gvn.transform( new AndINode(hsize, hsm) );
3709 Node* mask = intcon(round_mask);
3710 header_size = _gvn.transform( new AddINode(hsize, mask) );
3711 }
3712
3713 Node* elem_shift = NULL;
3714 if (layout_is_con) {
3715 int eshift = Klass::layout_helper_log2_element_size(layout_con);
3716 if (eshift != 0)
3717 elem_shift = intcon(eshift);
3718 } else {
3719 // There is no need to mask or shift this value.
3720 // The semantics of LShiftINode include an implicit mask to 0x1F.
3764 // places, one where the length is sharply limited, and the other
3765 // after a successful allocation.
3766 Node* abody = lengthx;
3767 if (elem_shift != NULL)
3768 abody = _gvn.transform( new LShiftXNode(lengthx, elem_shift) );
3769 Node* size = _gvn.transform( new AddXNode(headerx, abody) );
3770 if (round_mask != 0) {
3771 Node* mask = MakeConX(~round_mask);
3772 size = _gvn.transform( new AndXNode(size, mask) );
3773 }
3774 // else if round_mask == 0, the size computation is self-rounding
3775
3776 if (return_size_val != NULL) {
3777 // This is the size
3778 (*return_size_val) = size;
3779 }
3780
3781 // Now generate allocation code
3782
3783 // The entire memory state is needed for slow path of the allocation
3784 // since GC and deoptimization can happened.
3785 Node *mem = reset_memory();
3786 set_all_memory(mem); // Create new memory state
3787
3788 if (initial_slow_test->is_Bool()) {
3789 // Hide it behind a CMoveI, or else PhaseIdealLoop::split_up will get sick.
3790 initial_slow_test = initial_slow_test->as_Bool()->as_int_value(&_gvn);
3791 }
3792
3793 // Create the AllocateArrayNode and its result projections
3794 AllocateArrayNode* alloc
3795 = new AllocateArrayNode(C, AllocateArrayNode::alloc_type(TypeInt::INT),
3796 control(), mem, i_o(),
3797 size, klass_node,
3798 initial_slow_test,
3799 length);
3800
3801 // Cast to correct type. Note that the klass_node may be constant or not,
3802 // and in the latter case the actual array type will be inexact also.
3803 // (This happens via a non-constant argument to inline_native_newArray.)
3804 // In any case, the value of klass_node provides the desired array type.
3805 const TypeInt* length_type = _gvn.find_int_type(length);
3806 const TypeOopPtr* ary_type = _gvn.type(klass_node)->is_klassptr()->as_instance_type();
3807 if (ary_type->isa_aryptr() && length_type != NULL) {
3808 // Try to get a better type than POS for the size
3809 ary_type = ary_type->is_aryptr()->cast_to_size(length_type);
3810 }
3811
3812 Node* javaoop = set_output_for_allocation(alloc, ary_type, deoptimize_on_exception);
3813
3814 // Cast length on remaining path to be as narrow as possible
3815 if (map()->find_edge(length) >= 0) {
3816 Node* ccast = alloc->make_ideal_length(ary_type, &_gvn);
3817 if (ccast != length) {
3818 _gvn.set_type_bottom(ccast);
3819 record_for_igvn(ccast);
3820 replace_in_map(length, ccast);
3821 }
3822 }
3823
3824 return javaoop;
3825 }
3826
3941 set_all_memory(ideal.merged_memory());
3942 set_i_o(ideal.i_o());
3943 set_control(ideal.ctrl());
3944 }
3945
3946 void GraphKit::final_sync(IdealKit& ideal) {
3947 // Final sync IdealKit and graphKit.
3948 sync_kit(ideal);
3949 }
3950
3951 Node* GraphKit::load_String_length(Node* str, bool set_ctrl) {
3952 Node* len = load_array_length(load_String_value(str, set_ctrl));
3953 Node* coder = load_String_coder(str, set_ctrl);
3954 // Divide length by 2 if coder is UTF16
3955 return _gvn.transform(new RShiftINode(len, coder));
3956 }
3957
3958 Node* GraphKit::load_String_value(Node* str, bool set_ctrl) {
3959 int value_offset = java_lang_String::value_offset_in_bytes();
3960 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
3961 false, NULL, 0);
3962 const TypePtr* value_field_type = string_type->add_offset(value_offset);
3963 const TypeAryPtr* value_type = TypeAryPtr::make(TypePtr::NotNull,
3964 TypeAry::make(TypeInt::BYTE, TypeInt::POS),
3965 ciTypeArrayKlass::make(T_BYTE), true, 0);
3966 Node* p = basic_plus_adr(str, str, value_offset);
3967 Node* load = access_load_at(str, p, value_field_type, value_type, T_OBJECT,
3968 IN_HEAP | (set_ctrl ? C2_CONTROL_DEPENDENT_LOAD : 0) | MO_UNORDERED);
3969 return load;
3970 }
3971
3972 Node* GraphKit::load_String_coder(Node* str, bool set_ctrl) {
3973 if (!CompactStrings) {
3974 return intcon(java_lang_String::CODER_UTF16);
3975 }
3976 int coder_offset = java_lang_String::coder_offset_in_bytes();
3977 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
3978 false, NULL, 0);
3979 const TypePtr* coder_field_type = string_type->add_offset(coder_offset);
3980
3981 Node* p = basic_plus_adr(str, str, coder_offset);
3982 Node* load = access_load_at(str, p, coder_field_type, TypeInt::BYTE, T_BYTE,
3983 IN_HEAP | (set_ctrl ? C2_CONTROL_DEPENDENT_LOAD : 0) | MO_UNORDERED);
3984 return load;
3985 }
3986
3987 void GraphKit::store_String_value(Node* str, Node* value) {
3988 int value_offset = java_lang_String::value_offset_in_bytes();
3989 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
3990 false, NULL, 0);
3991 const TypePtr* value_field_type = string_type->add_offset(value_offset);
3992
3993 access_store_at(str, basic_plus_adr(str, value_offset), value_field_type,
3994 value, TypeAryPtr::BYTES, T_OBJECT, IN_HEAP | MO_UNORDERED);
3995 }
3996
3997 void GraphKit::store_String_coder(Node* str, Node* value) {
3998 int coder_offset = java_lang_String::coder_offset_in_bytes();
3999 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4000 false, NULL, 0);
4001 const TypePtr* coder_field_type = string_type->add_offset(coder_offset);
4002
4003 access_store_at(str, basic_plus_adr(str, coder_offset), coder_field_type,
4004 value, TypeInt::BYTE, T_BYTE, IN_HEAP | MO_UNORDERED);
4005 }
4006
4007 // Capture src and dst memory state with a MergeMemNode
4008 Node* GraphKit::capture_memory(const TypePtr* src_type, const TypePtr* dst_type) {
4009 if (src_type == dst_type) {
4010 // Types are equal, we don't need a MergeMemNode
4011 return memory(src_type);
4012 }
4013 MergeMemNode* merge = MergeMemNode::make(map()->memory());
4014 record_for_igvn(merge); // fold it up later, if possible
4015 int src_idx = C->get_alias_index(src_type);
4016 int dst_idx = C->get_alias_index(dst_type);
4017 merge->set_memory_at(src_idx, memory(src_idx));
4018 merge->set_memory_at(dst_idx, memory(dst_idx));
4019 return merge;
4020 }
4093 i_char->init_req(2, AddI(i_char, intcon(2)));
4094
4095 set_control(IfFalse(iff));
4096 set_memory(st, TypeAryPtr::BYTES);
4097 }
4098
4099 Node* GraphKit::make_constant_from_field(ciField* field, Node* obj) {
4100 if (!field->is_constant()) {
4101 return NULL; // Field not marked as constant.
4102 }
4103 ciInstance* holder = NULL;
4104 if (!field->is_static()) {
4105 ciObject* const_oop = obj->bottom_type()->is_oopptr()->const_oop();
4106 if (const_oop != NULL && const_oop->is_instance()) {
4107 holder = const_oop->as_instance();
4108 }
4109 }
4110 const Type* con_type = Type::make_constant_from_field(field, holder, field->layout_type(),
4111 /*is_unsigned_load=*/false);
4112 if (con_type != NULL) {
4113 return makecon(con_type);
4114 }
4115 return NULL;
4116 }
|
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "precompiled.hpp"
26 #include "ci/ciUtilities.hpp"
27 #include "compiler/compileLog.hpp"
28 #include "ci/ciValueKlass.hpp"
29 #include "gc/shared/barrierSet.hpp"
30 #include "gc/shared/c2/barrierSetC2.hpp"
31 #include "interpreter/interpreter.hpp"
32 #include "memory/resourceArea.hpp"
33 #include "opto/addnode.hpp"
34 #include "opto/castnode.hpp"
35 #include "opto/convertnode.hpp"
36 #include "opto/graphKit.hpp"
37 #include "opto/idealKit.hpp"
38 #include "opto/intrinsicnode.hpp"
39 #include "opto/locknode.hpp"
40 #include "opto/machnode.hpp"
41 #include "opto/narrowptrnode.hpp"
42 #include "opto/opaquenode.hpp"
43 #include "opto/parse.hpp"
44 #include "opto/rootnode.hpp"
45 #include "opto/runtime.hpp"
46 #include "opto/valuetypenode.hpp"
47 #include "runtime/deoptimization.hpp"
48 #include "runtime/sharedRuntime.hpp"
49
50 //----------------------------GraphKit-----------------------------------------
51 // Main utility constructor.
52 GraphKit::GraphKit(JVMState* jvms, PhaseGVN* gvn)
53 : Phase(Phase::Parser),
54 _env(C->env()),
55 _gvn((gvn != NULL) ? *gvn : *C->initial_gvn()),
56 _barrier_set(BarrierSet::barrier_set()->barrier_set_c2())
57 {
58 assert(gvn == NULL || !gvn->is_IterGVN() || gvn->is_IterGVN()->delay_transform(), "delay transform should be enabled");
59 _exceptions = jvms->map()->next_exception();
60 if (_exceptions != NULL) jvms->map()->set_next_exception(NULL);
61 set_jvms(jvms);
62 #ifdef ASSERT
63 if (_gvn.is_IterGVN() != NULL) {
64 assert(_gvn.is_IterGVN()->delay_transform(), "Transformation must be delayed if IterGVN is used");
65 // Save the initial size of _for_igvn worklist for verification (see ~GraphKit)
66 _worklist_size = _gvn.C->for_igvn()->size();
67 }
68 #endif
69 }
70
71 // Private constructor for parser.
72 GraphKit::GraphKit()
73 : Phase(Phase::Parser),
74 _env(C->env()),
75 _gvn(*C->initial_gvn()),
76 _barrier_set(BarrierSet::barrier_set()->barrier_set_c2())
77 {
78 _exceptions = NULL;
79 set_map(NULL);
80 debug_only(_sp = -99);
81 debug_only(set_bci(-99));
82 }
83
84
85
86 //---------------------------clean_stack---------------------------------------
87 // Clear away rubbish from the stack area of the JVM state.
88 // This destroys any arguments that may be waiting on the stack.
817 tty->print_cr("Zombie local %d: ", local);
818 jvms->dump();
819 }
820 return false;
821 }
822 }
823 }
824 return true;
825 }
826
827 #endif //ASSERT
828
829 // Helper function for enforcing certain bytecodes to reexecute if
830 // deoptimization happens
831 static bool should_reexecute_implied_by_bytecode(JVMState *jvms, bool is_anewarray) {
832 ciMethod* cur_method = jvms->method();
833 int cur_bci = jvms->bci();
834 if (cur_method != NULL && cur_bci != InvocationEntryBci) {
835 Bytecodes::Code code = cur_method->java_code_at_bci(cur_bci);
836 return Interpreter::bytecode_should_reexecute(code) ||
837 (is_anewarray && (code == Bytecodes::_multianewarray));
838 // Reexecute _multianewarray bytecode which was replaced with
839 // sequence of [a]newarray. See Parse::do_multianewarray().
840 //
841 // Note: interpreter should not have it set since this optimization
842 // is limited by dimensions and guarded by flag so in some cases
843 // multianewarray() runtime calls will be generated and
844 // the bytecode should not be reexecutes (stack will not be reset).
845 } else {
846 return false;
847 }
848 }
849
850 // Helper function for adding JVMState and debug information to node
851 void GraphKit::add_safepoint_edges(SafePointNode* call, bool must_throw) {
852 // Add the safepoint edges to the call (or other safepoint).
853
854 // Make sure dead locals are set to top. This
855 // should help register allocation time and cut down on the size
856 // of the deoptimization information.
857 assert(dead_locals_are_killed(), "garbage in debug info before safepoint");
858
859 // Walk the inline list to fill in the correct set of JVMState's
860 // Also fill in the associated edges for each JVMState.
861
862 // If the bytecode needs to be reexecuted we need to put
863 // the arguments back on the stack.
864 const bool should_reexecute = jvms()->should_reexecute();
865 JVMState* youngest_jvms = should_reexecute ? sync_jvms_for_reexecute() : sync_jvms();
866
867 // NOTE: set_bci (called from sync_jvms) might reset the reexecute bit to
1071 ciSignature* declared_signature = NULL;
1072 ciMethod* ignored_callee = method()->get_method_at_bci(bci(), ignored_will_link, &declared_signature);
1073 assert(declared_signature != NULL, "cannot be null");
1074 inputs = declared_signature->arg_size_for_bc(code);
1075 int size = declared_signature->return_type()->size();
1076 depth = size - inputs;
1077 }
1078 break;
1079
1080 case Bytecodes::_multianewarray:
1081 {
1082 ciBytecodeStream iter(method());
1083 iter.reset_to_bci(bci());
1084 iter.next();
1085 inputs = iter.get_dimensions();
1086 assert(rsize == 1, "");
1087 depth = rsize - inputs;
1088 }
1089 break;
1090
1091 case Bytecodes::_withfield: {
1092 bool ignored_will_link;
1093 ciField* field = method()->get_field_at_bci(bci(), ignored_will_link);
1094 int size = field->type()->size();
1095 inputs = size+1;
1096 depth = rsize - inputs;
1097 break;
1098 }
1099
1100 case Bytecodes::_ireturn:
1101 case Bytecodes::_lreturn:
1102 case Bytecodes::_freturn:
1103 case Bytecodes::_dreturn:
1104 case Bytecodes::_areturn:
1105 assert(rsize == -depth, "");
1106 inputs = rsize;
1107 break;
1108
1109 case Bytecodes::_jsr:
1110 case Bytecodes::_jsr_w:
1111 inputs = 0;
1112 depth = 1; // S.B. depth=1, not zero
1113 break;
1114
1115 default:
1116 // bytecode produces a typed result
1117 inputs = rsize - depth;
1118 assert(inputs >= 0, "");
1119 break;
1205 // the incoming address with NULL casted away. You are allowed to use the
1206 // not-null value only if you are control dependent on the test.
1207 #ifndef PRODUCT
1208 extern int explicit_null_checks_inserted,
1209 explicit_null_checks_elided;
1210 #endif
1211 Node* GraphKit::null_check_common(Node* value, BasicType type,
1212 // optional arguments for variations:
1213 bool assert_null,
1214 Node* *null_control,
1215 bool speculative) {
1216 assert(!assert_null || null_control == NULL, "not both at once");
1217 if (stopped()) return top();
1218 NOT_PRODUCT(explicit_null_checks_inserted++);
1219
1220 // Construct NULL check
1221 Node *chk = NULL;
1222 switch(type) {
1223 case T_LONG : chk = new CmpLNode(value, _gvn.zerocon(T_LONG)); break;
1224 case T_INT : chk = new CmpINode(value, _gvn.intcon(0)); break;
1225 case T_VALUETYPE : // fall through
1226 case T_ARRAY : // fall through
1227 type = T_OBJECT; // simplify further tests
1228 case T_OBJECT : {
1229 const Type *t = _gvn.type( value );
1230
1231 const TypeOopPtr* tp = t->isa_oopptr();
1232 if (tp != NULL && tp->klass() != NULL && !tp->klass()->is_loaded()
1233 // Only for do_null_check, not any of its siblings:
1234 && !assert_null && null_control == NULL) {
1235 // Usually, any field access or invocation on an unloaded oop type
1236 // will simply fail to link, since the statically linked class is
1237 // likely also to be unloaded. However, in -Xcomp mode, sometimes
1238 // the static class is loaded but the sharper oop type is not.
1239 // Rather than checking for this obscure case in lots of places,
1240 // we simply observe that a null check on an unloaded class
1241 // will always be followed by a nonsense operation, so we
1242 // can just issue the uncommon trap here.
1243 // Our access to the unloaded class will only be correct
1244 // after it has been loaded and initialized, which requires
1245 // a trip through the interpreter.
1377 }
1378
1379 if (assert_null) {
1380 // Cast obj to null on this path.
1381 replace_in_map(value, zerocon(type));
1382 return zerocon(type);
1383 }
1384
1385 // Cast obj to not-null on this path, if there is no null_control.
1386 // (If there is a null_control, a non-null value may come back to haunt us.)
1387 if (type == T_OBJECT) {
1388 Node* cast = cast_not_null(value, false);
1389 if (null_control == NULL || (*null_control) == top())
1390 replace_in_map(value, cast);
1391 value = cast;
1392 }
1393
1394 return value;
1395 }
1396
1397 Node* GraphKit::null2default(Node* value, ciValueKlass* vk) {
1398 Node* null_ctl = top();
1399 value = null_check_oop(value, &null_ctl);
1400 if (!null_ctl->is_top()) {
1401 // Return default value if oop is null
1402 Node* region = new RegionNode(3);
1403 region->init_req(1, control());
1404 region->init_req(2, null_ctl);
1405 value = PhiNode::make(region, value, TypeInstPtr::make(TypePtr::BotPTR, vk));
1406 value->set_req(2, ValueTypeNode::default_oop(gvn(), vk));
1407 set_control(gvn().transform(region));
1408 value = gvn().transform(value);
1409 }
1410 return value;
1411 }
1412
1413 //------------------------------cast_not_null----------------------------------
1414 // Cast obj to not-null on this path
1415 Node* GraphKit::cast_not_null(Node* obj, bool do_replace_in_map) {
1416 if (obj->is_ValueType()) {
1417 return obj;
1418 }
1419 const Type *t = _gvn.type(obj);
1420 const Type *t_not_null = t->join_speculative(TypePtr::NOTNULL);
1421 // Object is already not-null?
1422 if( t == t_not_null ) return obj;
1423
1424 Node *cast = new CastPPNode(obj,t_not_null);
1425 cast->init_req(0, control());
1426 cast = _gvn.transform( cast );
1427
1428 // Scan for instances of 'obj' in the current JVM mapping.
1429 // These instances are known to be not-null after the test.
1430 if (do_replace_in_map)
1431 replace_in_map(obj, cast);
1432
1433 return cast; // Return casted value
1434 }
1435
1436 // Sometimes in intrinsics, we implicitly know an object is not null
1437 // (there's no actual null check) so we can cast it to not null. In
1438 // the course of optimizations, the input to the cast can become null.
1527 int adr_idx,
1528 MemNode::MemOrd mo,
1529 LoadNode::ControlDependency control_dependency,
1530 bool require_atomic_access,
1531 bool unaligned,
1532 bool mismatched,
1533 bool unsafe) {
1534 assert(adr_idx != Compile::AliasIdxTop, "use other make_load factory" );
1535 const TypePtr* adr_type = NULL; // debug-mode-only argument
1536 debug_only(adr_type = C->get_adr_type(adr_idx));
1537 Node* mem = memory(adr_idx);
1538 Node* ld;
1539 if (require_atomic_access && bt == T_LONG) {
1540 ld = LoadLNode::make_atomic(ctl, mem, adr, adr_type, t, mo, control_dependency, unaligned, mismatched, unsafe);
1541 } else if (require_atomic_access && bt == T_DOUBLE) {
1542 ld = LoadDNode::make_atomic(ctl, mem, adr, adr_type, t, mo, control_dependency, unaligned, mismatched, unsafe);
1543 } else {
1544 ld = LoadNode::make(_gvn, ctl, mem, adr, adr_type, t, bt, mo, control_dependency, unaligned, mismatched, unsafe);
1545 }
1546 ld = _gvn.transform(ld);
1547
1548 if (((bt == T_OBJECT || bt == T_VALUETYPE) && C->do_escape_analysis()) || C->eliminate_boxing()) {
1549 // Improve graph before escape analysis and boxing elimination.
1550 record_for_igvn(ld);
1551 }
1552 return ld;
1553 }
1554
1555 Node* GraphKit::store_to_memory(Node* ctl, Node* adr, Node *val, BasicType bt,
1556 int adr_idx,
1557 MemNode::MemOrd mo,
1558 bool require_atomic_access,
1559 bool unaligned,
1560 bool mismatched,
1561 bool unsafe) {
1562 assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory" );
1563 const TypePtr* adr_type = NULL;
1564 debug_only(adr_type = C->get_adr_type(adr_idx));
1565 Node *mem = memory(adr_idx);
1566 Node* st;
1567 if (require_atomic_access && bt == T_LONG) {
1568 st = StoreLNode::make_atomic(ctl, mem, adr, adr_type, val, mo);
1579 }
1580 if (unsafe) {
1581 st->as_Store()->set_unsafe_access();
1582 }
1583 st = _gvn.transform(st);
1584 set_memory(st, adr_idx);
1585 // Back-to-back stores can only remove intermediate store with DU info
1586 // so push on worklist for optimizer.
1587 if (mem->req() > MemNode::Address && adr == mem->in(MemNode::Address))
1588 record_for_igvn(st);
1589
1590 return st;
1591 }
1592
1593 Node* GraphKit::access_store_at(Node* obj,
1594 Node* adr,
1595 const TypePtr* adr_type,
1596 Node* val,
1597 const Type* val_type,
1598 BasicType bt,
1599 DecoratorSet decorators,
1600 bool deoptimize_on_exception,
1601 bool safe_for_replace) {
1602 // Transformation of a value which could be NULL pointer (CastPP #NULL)
1603 // could be delayed during Parse (for example, in adjust_map_after_if()).
1604 // Execute transformation here to avoid barrier generation in such case.
1605 if (_gvn.type(val) == TypePtr::NULL_PTR) {
1606 val = _gvn.makecon(TypePtr::NULL_PTR);
1607 }
1608
1609 if (stopped()) {
1610 return top(); // Dead path ?
1611 }
1612
1613 assert(val != NULL, "not dead path");
1614 if (val->is_ValueType()) {
1615 // Allocate value type and get oop
1616 val = val->as_ValueType()->allocate(this, deoptimize_on_exception, safe_for_replace)->get_oop();
1617 }
1618
1619 C2AccessValuePtr addr(adr, adr_type);
1620 C2AccessValue value(val, val_type);
1621 C2ParseAccess access(this, decorators | C2_WRITE_ACCESS, bt, obj, addr);
1622 if (access.is_raw()) {
1623 return _barrier_set->BarrierSetC2::store_at(access, value);
1624 } else {
1625 return _barrier_set->store_at(access, value);
1626 }
1627 }
1628
1629 Node* GraphKit::access_load_at(Node* obj, // containing obj
1630 Node* adr, // actual adress to store val at
1631 const TypePtr* adr_type,
1632 const Type* val_type,
1633 BasicType bt,
1634 DecoratorSet decorators,
1635 Node* ctl) {
1636 if (stopped()) {
1637 return top(); // Dead path ?
1638 }
1639
1640 C2AccessValuePtr addr(adr, adr_type);
1641 C2ParseAccess access(this, decorators | C2_READ_ACCESS, bt, obj, addr, ctl);
1642 if (access.is_raw()) {
1643 return _barrier_set->BarrierSetC2::load_at(access, val_type);
1644 } else {
1645 return _barrier_set->load_at(access, val_type);
1646 }
1647 }
1648
1649 Node* GraphKit::access_load(Node* adr, // actual adress to load val at
1650 const Type* val_type,
1651 BasicType bt,
1652 DecoratorSet decorators) {
1653 if (stopped()) {
1654 return top(); // Dead path ?
1655 }
1656
1657 C2AccessValuePtr addr(adr, NULL);
1658 C2ParseAccess access(this, decorators | C2_READ_ACCESS, bt, NULL, addr);
1659 if (access.is_raw()) {
1660 return _barrier_set->BarrierSetC2::load_at(access, val_type);
1661 } else {
1719 }
1720 }
1721
1722 Node* GraphKit::access_atomic_add_at(Node* obj,
1723 Node* adr,
1724 const TypePtr* adr_type,
1725 int alias_idx,
1726 Node* new_val,
1727 const Type* value_type,
1728 BasicType bt,
1729 DecoratorSet decorators) {
1730 C2AccessValuePtr addr(adr, adr_type);
1731 C2AtomicParseAccess access(this, decorators | C2_READ_ACCESS | C2_WRITE_ACCESS, bt, obj, addr, alias_idx);
1732 if (access.is_raw()) {
1733 return _barrier_set->BarrierSetC2::atomic_add_at(access, new_val, value_type);
1734 } else {
1735 return _barrier_set->atomic_add_at(access, new_val, value_type);
1736 }
1737 }
1738
1739 void GraphKit::access_clone(Node* src_base, Node* dst_base, Node* countx, bool is_array) {
1740 return _barrier_set->clone(this, src_base, dst_base, countx, is_array);
1741 }
1742
1743 Node* GraphKit::access_resolve(Node* n, DecoratorSet decorators) {
1744 // Use stronger ACCESS_WRITE|ACCESS_READ by default.
1745 if ((decorators & (ACCESS_READ | ACCESS_WRITE)) == 0) {
1746 decorators |= ACCESS_READ | ACCESS_WRITE;
1747 }
1748 return _barrier_set->resolve(this, n, decorators);
1749 }
1750
1751 //-------------------------array_element_address-------------------------
1752 Node* GraphKit::array_element_address(Node* ary, Node* idx, BasicType elembt,
1753 const TypeInt* sizetype, Node* ctrl) {
1754 uint shift = exact_log2(type2aelembytes(elembt));
1755 ciKlass* arytype_klass = _gvn.type(ary)->is_aryptr()->klass();
1756 if (arytype_klass != NULL && arytype_klass->is_value_array_klass()) {
1757 ciValueArrayKlass* vak = arytype_klass->as_value_array_klass();
1758 shift = vak->log2_element_size();
1759 }
1760 uint header = arrayOopDesc::base_offset_in_bytes(elembt);
1761
1762 // short-circuit a common case (saves lots of confusing waste motion)
1763 jint idx_con = find_int_con(idx, -1);
1764 if (idx_con >= 0) {
1765 intptr_t offset = header + ((intptr_t)idx_con << shift);
1766 return basic_plus_adr(ary, offset);
1767 }
1768
1769 // must be correct type for alignment purposes
1770 Node* base = basic_plus_adr(ary, header);
1771 idx = Compile::conv_I2X_index(&_gvn, idx, sizetype, ctrl);
1772 Node* scale = _gvn.transform( new LShiftXNode(idx, intcon(shift)) );
1773 return basic_plus_adr(ary, base, scale);
1774 }
1775
1776 //-------------------------load_array_element-------------------------
1777 Node* GraphKit::load_array_element(Node* ctl, Node* ary, Node* idx, const TypeAryPtr* arytype) {
1778 const Type* elemtype = arytype->elem();
1779 BasicType elembt = elemtype->array_element_basic_type();
1780 assert(elembt != T_VALUETYPE, "value types are not supported by this method");
1781 Node* adr = array_element_address(ary, idx, elembt, arytype->size());
1782 if (elembt == T_NARROWOOP) {
1783 elembt = T_OBJECT; // To satisfy switch in LoadNode::make()
1784 }
1785 Node* ld = make_load(ctl, adr, elemtype, elembt, arytype, MemNode::unordered);
1786 return ld;
1787 }
1788
1789 //-------------------------set_arguments_for_java_call-------------------------
1790 // Arguments (pre-popped from the stack) are taken from the JVMS.
1791 void GraphKit::set_arguments_for_java_call(CallJavaNode* call, bool incremental_inlining) {
1792 // Add the call arguments:
1793 const TypeTuple* domain = call->tf()->domain_sig();
1794 ExtendedSignature sig_cc = ExtendedSignature(call->method()->get_sig_cc(), SigEntryFilter());
1795 uint nargs = domain->cnt();
1796 for (uint i = TypeFunc::Parms, idx = TypeFunc::Parms; i < nargs; i++) {
1797 Node* arg = argument(i-TypeFunc::Parms);
1798 const Type* t = domain->field_at(i);
1799 if (call->method()->has_scalarized_args() && t->is_valuetypeptr() && !t->maybe_null()) {
1800 // We don't pass value type arguments by reference but instead pass each field of the value type
1801 ValueTypeNode* vt = arg->as_ValueType();
1802 vt->pass_fields(this, call, sig_cc, idx);
1803 // If a value type argument is passed as fields, attach the Method* to the call site
1804 // to be able to access the extended signature later via attached_method_before_pc().
1805 // For example, see CompiledMethod::preserve_callee_argument_oops().
1806 call->set_override_symbolic_info(true);
1807 continue;
1808 } else if (arg->is_ValueType()) {
1809 // Pass value type argument via oop to callee
1810 if (!incremental_inlining) {
1811 arg = arg->as_ValueType()->allocate(this)->get_oop();
1812 } else {
1813 arg = ValueTypePtrNode::make_from_value_type(this, arg->as_ValueType());
1814 }
1815 }
1816 call->init_req(idx++, arg);
1817 // Skip reserved arguments
1818 BasicType bt = t->basic_type();
1819 while (SigEntry::next_is_reserved(sig_cc, bt, true)) {
1820 call->init_req(idx++, top());
1821 if (type2size[bt] == 2) {
1822 call->init_req(idx++, top());
1823 }
1824 }
1825 }
1826 }
1827
1828 //---------------------------set_edges_for_java_call---------------------------
1829 // Connect a newly created call into the current JVMS.
1830 // A return value node (if any) is returned from set_edges_for_java_call.
1831 void GraphKit::set_edges_for_java_call(CallJavaNode* call, bool must_throw, bool separate_io_proj) {
1832
1833 // Add the predefined inputs:
1834 call->init_req( TypeFunc::Control, control() );
1835 call->init_req( TypeFunc::I_O , i_o() );
1836 call->init_req( TypeFunc::Memory , reset_memory() );
1837 call->init_req( TypeFunc::FramePtr, frameptr() );
1838 call->init_req( TypeFunc::ReturnAdr, top() );
1839
1840 add_safepoint_edges(call, must_throw);
1841
1842 Node* xcall = _gvn.transform(call);
1843
1844 if (xcall == top()) {
1845 set_control(top());
1846 return;
1847 }
1848 assert(xcall == call, "call identity is stable");
1849
1850 // Re-use the current map to produce the result.
1851
1852 set_control(_gvn.transform(new ProjNode(call, TypeFunc::Control)));
1853 set_i_o( _gvn.transform(new ProjNode(call, TypeFunc::I_O , separate_io_proj)));
1854 set_all_memory_call(xcall, separate_io_proj);
1855
1856 //return xcall; // no need, caller already has it
1857 }
1858
1859 Node* GraphKit::set_results_for_java_call(CallJavaNode* call, bool separate_io_proj, bool deoptimize) {
1860 if (stopped()) return top(); // maybe the call folded up?
1861
1862 // Note: Since any out-of-line call can produce an exception,
1863 // we always insert an I_O projection from the call into the result.
1864
1865 make_slow_call_ex(call, env()->Throwable_klass(), separate_io_proj, deoptimize);
1866
1867 if (separate_io_proj) {
1868 // The caller requested separate projections be used by the fall
1869 // through and exceptional paths, so replace the projections for
1870 // the fall through path.
1871 set_i_o(_gvn.transform( new ProjNode(call, TypeFunc::I_O) ));
1872 set_all_memory(_gvn.transform( new ProjNode(call, TypeFunc::Memory) ));
1873 }
1874
1875 // Capture the return value, if any.
1876 Node* ret;
1877 if (call->method() == NULL || call->method()->return_type()->basic_type() == T_VOID) {
1878 ret = top();
1879 } else if (call->tf()->returns_value_type_as_fields()) {
1880 // Return of multiple values (value type fields): we create a
1881 // ValueType node, each field is a projection from the call.
1882 ciValueKlass* vk = call->method()->return_type()->as_value_klass();
1883 const Array<SigEntry>* sig_array = vk->extended_sig();
1884 GrowableArray<SigEntry> sig = GrowableArray<SigEntry>(sig_array->length());
1885 sig.appendAll(sig_array);
1886 ExtendedSignature sig_cc = ExtendedSignature(&sig, SigEntryFilter());
1887 uint base_input = TypeFunc::Parms + 1;
1888 ret = ValueTypeNode::make_from_multi(this, call, sig_cc, vk, base_input, false);
1889 } else {
1890 ret = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
1891 }
1892
1893 return ret;
1894 }
1895
1896 //--------------------set_predefined_input_for_runtime_call--------------------
1897 // Reading and setting the memory state is way conservative here.
1898 // The real problem is that I am not doing real Type analysis on memory,
1899 // so I cannot distinguish card mark stores from other stores. Across a GC
1900 // point the Store Barrier and the card mark memory has to agree. I cannot
1901 // have a card mark store and its barrier split across the GC point from
1902 // either above or below. Here I get that to happen by reading ALL of memory.
1903 // A better answer would be to separate out card marks from other memory.
1904 // For now, return the input memory state, so that it can be reused
1905 // after the call, if this call has restricted memory effects.
1906 Node* GraphKit::set_predefined_input_for_runtime_call(SafePointNode* call, Node* narrow_mem) {
1907 // Set fixed predefined input arguments
1908 Node* memory = reset_memory();
1909 Node* m = narrow_mem == NULL ? memory : narrow_mem;
1910 call->init_req( TypeFunc::Control, control() );
1911 call->init_req( TypeFunc::I_O, top() ); // does no i/o
1912 call->init_req( TypeFunc::Memory, m ); // may gc ptrs
1963 if (use->is_MergeMem()) {
1964 wl.push(use);
1965 }
1966 }
1967 }
1968
1969 // Replace the call with the current state of the kit.
1970 void GraphKit::replace_call(CallNode* call, Node* result, bool do_replaced_nodes) {
1971 JVMState* ejvms = NULL;
1972 if (has_exceptions()) {
1973 ejvms = transfer_exceptions_into_jvms();
1974 }
1975
1976 ReplacedNodes replaced_nodes = map()->replaced_nodes();
1977 ReplacedNodes replaced_nodes_exception;
1978 Node* ex_ctl = top();
1979
1980 SafePointNode* final_state = stop();
1981
1982 // Find all the needed outputs of this call
1983 CallProjections* callprojs = call->extract_projections(true);
1984
1985 Unique_Node_List wl;
1986 Node* init_mem = call->in(TypeFunc::Memory);
1987 Node* final_mem = final_state->in(TypeFunc::Memory);
1988 Node* final_ctl = final_state->in(TypeFunc::Control);
1989 Node* final_io = final_state->in(TypeFunc::I_O);
1990
1991 // Replace all the old call edges with the edges from the inlining result
1992 if (callprojs->fallthrough_catchproj != NULL) {
1993 C->gvn_replace_by(callprojs->fallthrough_catchproj, final_ctl);
1994 }
1995 if (callprojs->fallthrough_memproj != NULL) {
1996 if (final_mem->is_MergeMem()) {
1997 // Parser's exits MergeMem was not transformed but may be optimized
1998 final_mem = _gvn.transform(final_mem);
1999 }
2000 C->gvn_replace_by(callprojs->fallthrough_memproj, final_mem);
2001 add_mergemem_users_to_worklist(wl, final_mem);
2002 }
2003 if (callprojs->fallthrough_ioproj != NULL) {
2004 C->gvn_replace_by(callprojs->fallthrough_ioproj, final_io);
2005 }
2006
2007 // Replace the result with the new result if it exists and is used
2008 if (callprojs->resproj[0] != NULL && result != NULL) {
2009 assert(callprojs->nb_resproj == 1, "unexpected number of results");
2010 C->gvn_replace_by(callprojs->resproj[0], result);
2011 }
2012
2013 if (ejvms == NULL) {
2014 // No exception edges to simply kill off those paths
2015 if (callprojs->catchall_catchproj != NULL) {
2016 C->gvn_replace_by(callprojs->catchall_catchproj, C->top());
2017 }
2018 if (callprojs->catchall_memproj != NULL) {
2019 C->gvn_replace_by(callprojs->catchall_memproj, C->top());
2020 }
2021 if (callprojs->catchall_ioproj != NULL) {
2022 C->gvn_replace_by(callprojs->catchall_ioproj, C->top());
2023 }
2024 // Replace the old exception object with top
2025 if (callprojs->exobj != NULL) {
2026 C->gvn_replace_by(callprojs->exobj, C->top());
2027 }
2028 } else {
2029 GraphKit ekit(ejvms);
2030
2031 // Load my combined exception state into the kit, with all phis transformed:
2032 SafePointNode* ex_map = ekit.combine_and_pop_all_exception_states();
2033 replaced_nodes_exception = ex_map->replaced_nodes();
2034
2035 Node* ex_oop = ekit.use_exception_state(ex_map);
2036
2037 if (callprojs->catchall_catchproj != NULL) {
2038 C->gvn_replace_by(callprojs->catchall_catchproj, ekit.control());
2039 ex_ctl = ekit.control();
2040 }
2041 if (callprojs->catchall_memproj != NULL) {
2042 Node* ex_mem = ekit.reset_memory();
2043 C->gvn_replace_by(callprojs->catchall_memproj, ex_mem);
2044 add_mergemem_users_to_worklist(wl, ex_mem);
2045 }
2046 if (callprojs->catchall_ioproj != NULL) {
2047 C->gvn_replace_by(callprojs->catchall_ioproj, ekit.i_o());
2048 }
2049
2050 // Replace the old exception object with the newly created one
2051 if (callprojs->exobj != NULL) {
2052 C->gvn_replace_by(callprojs->exobj, ex_oop);
2053 }
2054 }
2055
2056 // Disconnect the call from the graph
2057 call->disconnect_inputs(NULL, C);
2058 C->gvn_replace_by(call, C->top());
2059
2060 // Clean up any MergeMems that feed other MergeMems since the
2061 // optimizer doesn't like that.
2062 while (wl.size() > 0) {
2063 _gvn.transform(wl.pop());
2064 }
2065
2066 if (callprojs->fallthrough_catchproj != NULL && !final_ctl->is_top() && do_replaced_nodes) {
2067 replaced_nodes.apply(C, final_ctl);
2068 }
2069 if (!ex_ctl->is_top() && do_replaced_nodes) {
2070 replaced_nodes_exception.apply(C, ex_ctl);
2071 }
2072 }
2073
2074
2075 //------------------------------increment_counter------------------------------
2076 // for statistics: increment a VM counter by 1
2077
2078 void GraphKit::increment_counter(address counter_addr) {
2079 Node* adr1 = makecon(TypeRawPtr::make(counter_addr));
2080 increment_counter(adr1);
2081 }
2082
2083 void GraphKit::increment_counter(Node* counter_addr) {
2084 int adr_type = Compile::AliasIdxRaw;
2085 Node* ctrl = control();
2086 Node* cnt = make_load(ctrl, counter_addr, TypeInt::INT, T_INT, adr_type, MemNode::unordered);
2223 // it does not require card marks.
2224 Node* GraphKit::just_allocated_object(Node* current_control) {
2225 Node* ctrl = current_control;
2226 // Object::<init> is invoked after allocation, most of invoke nodes
2227 // will be reduced, but a region node is kept in parse time, we check
2228 // the pattern and skip the region node if it degraded to a copy.
2229 if (ctrl != NULL && ctrl->is_Region() && ctrl->req() == 2 &&
2230 ctrl->as_Region()->is_copy()) {
2231 ctrl = ctrl->as_Region()->is_copy();
2232 }
2233 if (C->recent_alloc_ctl() == ctrl) {
2234 return C->recent_alloc_obj();
2235 }
2236 return NULL;
2237 }
2238
2239
2240 void GraphKit::round_double_arguments(ciMethod* dest_method) {
2241 // (Note: TypeFunc::make has a cache that makes this fast.)
2242 const TypeFunc* tf = TypeFunc::make(dest_method);
2243 int nargs = tf->domain_sig()->cnt() - TypeFunc::Parms;
2244 for (int j = 0; j < nargs; j++) {
2245 const Type *targ = tf->domain_sig()->field_at(j + TypeFunc::Parms);
2246 if( targ->basic_type() == T_DOUBLE ) {
2247 // If any parameters are doubles, they must be rounded before
2248 // the call, dstore_rounding does gvn.transform
2249 Node *arg = argument(j);
2250 arg = dstore_rounding(arg);
2251 set_argument(j, arg);
2252 }
2253 }
2254 }
2255
2256 /**
2257 * Record profiling data exact_kls for Node n with the type system so
2258 * that it can propagate it (speculation)
2259 *
2260 * @param n node that the type applies to
2261 * @param exact_kls type from profiling
2262 * @param maybe_null did profiling see null?
2263 *
2264 * @return node with improved type
2265 */
2282 speculative = speculative->with_inline_depth(jvms()->depth());
2283 } else if (current_type->would_improve_ptr(ptr_kind)) {
2284 // Profiling report that null was never seen so we can change the
2285 // speculative type to non null ptr.
2286 if (ptr_kind == ProfileAlwaysNull) {
2287 speculative = TypePtr::NULL_PTR;
2288 } else {
2289 assert(ptr_kind == ProfileNeverNull, "nothing else is an improvement");
2290 const TypePtr* ptr = TypePtr::NOTNULL;
2291 if (speculative != NULL) {
2292 speculative = speculative->cast_to_ptr_type(ptr->ptr())->is_ptr();
2293 } else {
2294 speculative = ptr;
2295 }
2296 }
2297 }
2298
2299 if (speculative != current_type->speculative()) {
2300 // Build a type with a speculative type (what we think we know
2301 // about the type but will need a guard when we use it)
2302 const TypeOopPtr* spec_type = TypeOopPtr::make(TypePtr::BotPTR, Type::Offset::bottom, TypeOopPtr::InstanceBot, speculative);
2303 // We're changing the type, we need a new CheckCast node to carry
2304 // the new type. The new type depends on the control: what
2305 // profiling tells us is only valid from here as far as we can
2306 // tell.
2307 Node* cast = new CheckCastPPNode(control(), n, current_type->remove_speculative()->join_speculative(spec_type));
2308 cast = _gvn.transform(cast);
2309 replace_in_map(n, cast);
2310 n = cast;
2311 }
2312
2313 return n;
2314 }
2315
2316 /**
2317 * Record profiling data from receiver profiling at an invoke with the
2318 * type system so that it can propagate it (speculation)
2319 *
2320 * @param n receiver node
2321 *
2322 * @return node with improved type
2347 }
2348 ptr_kind = (i == call->row_limit()) ? ProfileAlwaysNull : ProfileMaybeNull;
2349 }
2350 }
2351 }
2352 return record_profile_for_speculation(n, exact_kls, ptr_kind);
2353 }
2354
2355 /**
2356 * Record profiling data from argument profiling at an invoke with the
2357 * type system so that it can propagate it (speculation)
2358 *
2359 * @param dest_method target method for the call
2360 * @param bc what invoke bytecode is this?
2361 */
2362 void GraphKit::record_profiled_arguments_for_speculation(ciMethod* dest_method, Bytecodes::Code bc) {
2363 if (!UseTypeSpeculation) {
2364 return;
2365 }
2366 const TypeFunc* tf = TypeFunc::make(dest_method);
2367 int nargs = tf->domain_sig()->cnt() - TypeFunc::Parms;
2368 int skip = Bytecodes::has_receiver(bc) ? 1 : 0;
2369 for (int j = skip, i = 0; j < nargs && i < TypeProfileArgsLimit; j++) {
2370 const Type *targ = tf->domain_sig()->field_at(j + TypeFunc::Parms);
2371 if (targ->isa_oopptr()) {
2372 ProfilePtrKind ptr_kind = ProfileMaybeNull;
2373 ciKlass* better_type = NULL;
2374 if (method()->argument_profiled_type(bci(), i, better_type, ptr_kind)) {
2375 record_profile_for_speculation(argument(j), better_type, ptr_kind);
2376 }
2377 i++;
2378 }
2379 }
2380 }
2381
2382 /**
2383 * Record profiling data from parameter profiling at an invoke with
2384 * the type system so that it can propagate it (speculation)
2385 */
2386 void GraphKit::record_profiled_parameters_for_speculation() {
2387 if (!UseTypeSpeculation) {
2388 return;
2389 }
2390 for (int i = 0, j = 0; i < method()->arg_size() ; i++) {
2391 if (_gvn.type(local(i))->isa_oopptr()) {
2866 // The decision to inline or out-of-line this final check is platform
2867 // dependent, and is found in the AD file definition of PartialSubtypeCheck.
2868 Node* psc = gvn->transform(
2869 new PartialSubtypeCheckNode(*ctrl, subklass, superklass));
2870
2871 IfNode *iff4 = gen_subtype_check_compare(*ctrl, psc, gvn->zerocon(T_OBJECT), BoolTest::ne, PROB_FAIR, gvn, T_ADDRESS);
2872 r_not_subtype->init_req(2, gvn->transform(new IfTrueNode (iff4)));
2873 r_ok_subtype ->init_req(3, gvn->transform(new IfFalseNode(iff4)));
2874
2875 // Return false path; set default control to true path.
2876 *ctrl = gvn->transform(r_ok_subtype);
2877 return gvn->transform(r_not_subtype);
2878 }
2879
2880 // Profile-driven exact type check:
2881 Node* GraphKit::type_check_receiver(Node* receiver, ciKlass* klass,
2882 float prob,
2883 Node* *casted_receiver) {
2884 const TypeKlassPtr* tklass = TypeKlassPtr::make(klass);
2885 Node* recv_klass = load_object_klass(receiver);
2886 Node* fail = type_check(recv_klass, tklass, prob);
2887 const TypeOopPtr* recv_xtype = tklass->as_instance_type();
2888 assert(recv_xtype->klass_is_exact(), "");
2889
2890 // Subsume downstream occurrences of receiver with a cast to
2891 // recv_xtype, since now we know what the type will be.
2892 Node* cast = new CheckCastPPNode(control(), receiver, recv_xtype);
2893 Node* res = _gvn.transform(cast);
2894 if (recv_xtype->is_valuetypeptr() && recv_xtype->value_klass()->is_scalarizable()) {
2895 assert(!gvn().type(res)->maybe_null(), "receiver should never be null");
2896 res = ValueTypeNode::make_from_oop(this, res, recv_xtype->value_klass());
2897 }
2898
2899 (*casted_receiver) = res;
2900 // (User must make the replace_in_map call.)
2901
2902 return fail;
2903 }
2904
2905 Node* GraphKit::type_check(Node* recv_klass, const TypeKlassPtr* tklass,
2906 float prob) {
2907 Node* want_klass = makecon(tklass);
2908 Node* cmp = _gvn.transform( new CmpPNode(recv_klass, want_klass));
2909 Node* bol = _gvn.transform( new BoolNode(cmp, BoolTest::eq) );
2910 IfNode* iff = create_and_xform_if(control(), bol, prob, COUNT_UNKNOWN);
2911 set_control( _gvn.transform( new IfTrueNode (iff)));
2912 Node* fail = _gvn.transform( new IfFalseNode(iff));
2913 return fail;
2914 }
2915
2916 //------------------------------subtype_check_receiver-------------------------
2917 Node* GraphKit::subtype_check_receiver(Node* receiver, ciKlass* klass,
2918 Node** casted_receiver) {
2919 const TypeKlassPtr* tklass = TypeKlassPtr::make(klass);
2920 Node* recv_klass = load_object_klass(receiver);
2921 Node* want_klass = makecon(tklass);
2922
2923 Node* slow_ctl = gen_subtype_check(recv_klass, want_klass);
2924
2925 // Cast receiver after successful check
2926 const TypeOopPtr* recv_type = tklass->cast_to_exactness(false)->is_klassptr()->as_instance_type();
2927 Node* cast = new CheckCastPPNode(control(), receiver, recv_type);
2928 (*casted_receiver) = _gvn.transform(cast);
2929
2930 return slow_ctl;
2931 }
2932
2933 //------------------------------seems_never_null-------------------------------
2934 // Use null_seen information if it is available from the profile.
2935 // If we see an unexpected null at a type check we record it and force a
3120 // and the reflective instance-of call.
3121 Node* GraphKit::gen_instanceof(Node* obj, Node* superklass, bool safe_for_replace) {
3122 kill_dead_locals(); // Benefit all the uncommon traps
3123 assert( !stopped(), "dead parse path should be checked in callers" );
3124 assert(!TypePtr::NULL_PTR->higher_equal(_gvn.type(superklass)->is_klassptr()),
3125 "must check for not-null not-dead klass in callers");
3126
3127 // Make the merge point
3128 enum { _obj_path = 1, _fail_path, _null_path, PATH_LIMIT };
3129 RegionNode* region = new RegionNode(PATH_LIMIT);
3130 Node* phi = new PhiNode(region, TypeInt::BOOL);
3131 C->set_has_split_ifs(true); // Has chance for split-if optimization
3132
3133 ciProfileData* data = NULL;
3134 if (java_bc() == Bytecodes::_instanceof) { // Only for the bytecode
3135 data = method()->method_data()->bci_to_data(bci());
3136 }
3137 bool speculative_not_null = false;
3138 bool never_see_null = (ProfileDynamicTypes // aggressive use of profile
3139 && seems_never_null(obj, data, speculative_not_null));
3140 bool is_value = obj->is_ValueType();
3141
3142 // Null check; get casted pointer; set region slot 3
3143 Node* null_ctl = top();
3144 Node* not_null_obj = is_value ? obj : null_check_oop(obj, &null_ctl, never_see_null, safe_for_replace, speculative_not_null);
3145
3146 // If not_null_obj is dead, only null-path is taken
3147 if (stopped()) { // Doing instance-of on a NULL?
3148 set_control(null_ctl);
3149 return intcon(0);
3150 }
3151 region->init_req(_null_path, null_ctl);
3152 phi ->init_req(_null_path, intcon(0)); // Set null path value
3153 if (null_ctl == top()) {
3154 // Do this eagerly, so that pattern matches like is_diamond_phi
3155 // will work even during parsing.
3156 assert(_null_path == PATH_LIMIT-1, "delete last");
3157 region->del_req(_null_path);
3158 phi ->del_req(_null_path);
3159 }
3160
3161 // Do we know the type check always succeed?
3162 if (!is_value) {
3163 bool known_statically = false;
3164 if (_gvn.type(superklass)->singleton()) {
3165 ciKlass* superk = _gvn.type(superklass)->is_klassptr()->klass();
3166 ciKlass* subk = _gvn.type(obj)->is_oopptr()->klass();
3167 if (subk != NULL && subk->is_loaded()) {
3168 int static_res = C->static_subtype_check(superk, subk);
3169 known_statically = (static_res == Compile::SSC_always_true || static_res == Compile::SSC_always_false);
3170 }
3171 }
3172
3173 if (!known_statically) {
3174 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr();
3175 // We may not have profiling here or it may not help us. If we
3176 // have a speculative type use it to perform an exact cast.
3177 ciKlass* spec_obj_type = obj_type->speculative_type();
3178 if (spec_obj_type != NULL || (ProfileDynamicTypes && data != NULL)) {
3179 Node* cast_obj = maybe_cast_profiled_receiver(not_null_obj, NULL, spec_obj_type, safe_for_replace);
3180 if (stopped()) { // Profile disagrees with this path.
3181 set_control(null_ctl); // Null is the only remaining possibility.
3182 return intcon(0);
3183 }
3184 if (cast_obj != NULL &&
3185 // A value that's sometimes null is not something we can optimize well
3186 !(cast_obj->is_ValueType() && null_ctl != top())) {
3187 not_null_obj = cast_obj;
3188 is_value = not_null_obj->is_ValueType();
3189 }
3190 }
3191 }
3192 }
3193
3194 // Load the object's klass
3195 Node* obj_klass = NULL;
3196 if (is_value) {
3197 obj_klass = makecon(TypeKlassPtr::make(_gvn.type(not_null_obj)->value_klass()));
3198 } else {
3199 obj_klass = load_object_klass(not_null_obj);
3200 }
3201
3202 // Generate the subtype check
3203 Node* not_subtype_ctrl = gen_subtype_check(obj_klass, superklass);
3204
3205 // Plug in the success path to the general merge in slot 1.
3206 region->init_req(_obj_path, control());
3207 phi ->init_req(_obj_path, intcon(1));
3208
3209 // Plug in the failing path to the general merge in slot 2.
3210 region->init_req(_fail_path, not_subtype_ctrl);
3211 phi ->init_req(_fail_path, intcon(0));
3212
3213 // Return final merged results
3214 set_control( _gvn.transform(region) );
3215 record_for_igvn(region);
3216
3217 // If we know the type check always succeeds then we don't use the
3218 // profiling data at this bytecode. Don't lose it, feed it to the
3219 // type system as a speculative type.
3220 if (safe_for_replace && !is_value) {
3221 Node* casted_obj = record_profiled_receiver_for_speculation(obj);
3222 replace_in_map(obj, casted_obj);
3223 }
3224
3225 return _gvn.transform(phi);
3226 }
3227
3228 //-------------------------------gen_checkcast---------------------------------
3229 // Generate a checkcast idiom. Used by both the checkcast bytecode and the
3230 // array store bytecode. Stack must be as-if BEFORE doing the bytecode so the
3231 // uncommon-trap paths work. Adjust stack after this call.
3232 // If failure_control is supplied and not null, it is filled in with
3233 // the control edge for the cast failure. Otherwise, an appropriate
3234 // uncommon trap or exception is thrown.
3235 Node* GraphKit::gen_checkcast(Node *obj, Node* superklass, Node* *failure_control, bool never_null) {
3236 kill_dead_locals(); // Benefit all the uncommon traps
3237 const TypeKlassPtr* tk = _gvn.type(superklass)->is_klassptr();
3238 const TypeOopPtr* toop = TypeOopPtr::make_from_klass(tk->klass());
3239 assert(!never_null || toop->is_valuetypeptr(), "must be a value type pointer");
3240 bool is_value = obj->is_ValueType();
3241
3242 // Fast cutout: Check the case that the cast is vacuously true.
3243 // This detects the common cases where the test will short-circuit
3244 // away completely. We do this before we perform the null check,
3245 // because if the test is going to turn into zero code, we don't
3246 // want a residual null check left around. (Causes a slowdown,
3247 // for example, in some objArray manipulations, such as a[i]=a[j].)
3248 if (tk->singleton()) {
3249 ciKlass* klass = NULL;
3250 if (is_value) {
3251 klass = _gvn.type(obj)->value_klass();
3252 } else {
3253 const TypeOopPtr* objtp = _gvn.type(obj)->isa_oopptr();
3254 if (objtp != NULL) {
3255 klass = objtp->klass();
3256 }
3257 }
3258 if (klass != NULL) {
3259 switch (C->static_subtype_check(tk->klass(), klass)) {
3260 case Compile::SSC_always_true:
3261 // If we know the type check always succeed then we don't use
3262 // the profiling data at this bytecode. Don't lose it, feed it
3263 // to the type system as a speculative type.
3264 if (!is_value) {
3265 obj = record_profiled_receiver_for_speculation(obj);
3266 if (never_null) {
3267 obj = null_check(obj);
3268 }
3269 if (toop->is_valuetypeptr() && toop->value_klass()->is_scalarizable() && !gvn().type(obj)->maybe_null()) {
3270 obj = ValueTypeNode::make_from_oop(this, obj, toop->value_klass());
3271 }
3272 }
3273 return obj;
3274 case Compile::SSC_always_false:
3275 if (is_value || never_null) {
3276 if (!is_value) {
3277 null_check(obj);
3278 }
3279 // Value type is never null. Always throw an exception.
3280 builtin_throw(Deoptimization::Reason_class_check, makecon(TypeKlassPtr::make(klass)));
3281 return top();
3282 } else {
3283 // It needs a null check because a null will *pass* the cast check.
3284 return null_assert(obj);
3285 }
3286 }
3287 }
3288 }
3289
3290 ciProfileData* data = NULL;
3291 bool safe_for_replace = false;
3292 if (failure_control == NULL) { // use MDO in regular case only
3293 assert(java_bc() == Bytecodes::_aastore ||
3294 java_bc() == Bytecodes::_checkcast,
3295 "interpreter profiles type checks only for these BCs");
3296 data = method()->method_data()->bci_to_data(bci());
3297 safe_for_replace = true;
3298 }
3299
3300 // Make the merge point
3301 enum { _obj_path = 1, _null_path, PATH_LIMIT };
3302 RegionNode* region = new RegionNode(PATH_LIMIT);
3303 Node* phi = new PhiNode(region, toop);
3304 _gvn.set_type(region, Type::CONTROL);
3305 _gvn.set_type(phi, toop);
3306
3307 C->set_has_split_ifs(true); // Has chance for split-if optimization
3308
3309 // Use null-cast information if it is available
3310 bool speculative_not_null = false;
3311 bool never_see_null = ((failure_control == NULL) // regular case only
3312 && seems_never_null(obj, data, speculative_not_null));
3313
3314 // Null check; get casted pointer; set region slot 3
3315 Node* null_ctl = top();
3316 Node* not_null_obj = NULL;
3317 if (is_value) {
3318 not_null_obj = obj;
3319 } else if (never_null) {
3320 not_null_obj = null_check(obj);
3321 } else {
3322 not_null_obj = null_check_oop(obj, &null_ctl, never_see_null, safe_for_replace, speculative_not_null);
3323 }
3324
3325 // If not_null_obj is dead, only null-path is taken
3326 if (stopped()) { // Doing instance-of on a NULL?
3327 set_control(null_ctl);
3328 return null();
3329 }
3330 region->init_req(_null_path, null_ctl);
3331 phi ->init_req(_null_path, null()); // Set null path value
3332 if (null_ctl == top()) {
3333 // Do this eagerly, so that pattern matches like is_diamond_phi
3334 // will work even during parsing.
3335 assert(_null_path == PATH_LIMIT-1, "delete last");
3336 region->del_req(_null_path);
3337 phi ->del_req(_null_path);
3338 }
3339
3340 Node* cast_obj = NULL;
3341 if (!is_value && tk->klass_is_exact()) {
3342 // The following optimization tries to statically cast the speculative type of the object
3343 // (for example obtained during profiling) to the type of the superklass and then do a
3344 // dynamic check that the type of the object is what we expect. To work correctly
3345 // for checkcast and aastore the type of superklass should be exact.
3346 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr();
3347 // We may not have profiling here or it may not help us. If we have
3348 // a speculative type use it to perform an exact cast.
3349 ciKlass* spec_obj_type = obj_type->speculative_type();
3350 if (spec_obj_type != NULL || data != NULL) {
3351 cast_obj = maybe_cast_profiled_receiver(not_null_obj, tk->klass(), spec_obj_type, safe_for_replace);
3352 if (cast_obj != NULL && cast_obj->is_ValueType()) {
3353 if (null_ctl != top()) {
3354 cast_obj = NULL; // A value that's sometimes null is not something we can optimize well
3355 } else {
3356 return cast_obj;
3357 }
3358 }
3359 if (cast_obj != NULL) {
3360 if (failure_control != NULL) // failure is now impossible
3361 (*failure_control) = top();
3362 // adjust the type of the phi to the exact klass:
3363 phi->raise_bottom_type(_gvn.type(cast_obj)->meet_speculative(TypePtr::NULL_PTR));
3364 }
3365 }
3366 }
3367
3368 if (cast_obj == NULL) {
3369 // Load the object's klass
3370 Node* obj_klass = NULL;
3371 if (is_value) {
3372 obj_klass = makecon(TypeKlassPtr::make(_gvn.type(not_null_obj)->value_klass()));
3373 } else {
3374 obj_klass = load_object_klass(not_null_obj);
3375 }
3376
3377 // Generate the subtype check
3378 Node* not_subtype_ctrl = gen_subtype_check( obj_klass, superklass );
3379
3380 // Plug in success path into the merge
3381 cast_obj = is_value ? not_null_obj : _gvn.transform(new CheckCastPPNode(control(), not_null_obj, toop));
3382 // Failure path ends in uncommon trap (or may be dead - failure impossible)
3383 if (failure_control == NULL) {
3384 if (not_subtype_ctrl != top()) { // If failure is possible
3385 PreserveJVMState pjvms(this);
3386 set_control(not_subtype_ctrl);
3387 builtin_throw(Deoptimization::Reason_class_check, obj_klass);
3388 }
3389 } else {
3390 (*failure_control) = not_subtype_ctrl;
3391 }
3392 }
3393
3394 region->init_req(_obj_path, control());
3395 phi ->init_req(_obj_path, cast_obj);
3396
3397 // A merge of NULL or Casted-NotNull obj
3398 Node* res = _gvn.transform(phi);
3399
3400 // Note I do NOT always 'replace_in_map(obj,result)' here.
3401 // if( tk->klass()->can_be_primary_super() )
3402 // This means that if I successfully store an Object into an array-of-String
3403 // I 'forget' that the Object is really now known to be a String. I have to
3404 // do this because we don't have true union types for interfaces - if I store
3405 // a Baz into an array-of-Interface and then tell the optimizer it's an
3406 // Interface, I forget that it's also a Baz and cannot do Baz-like field
3407 // references to it. FIX THIS WHEN UNION TYPES APPEAR!
3408 // replace_in_map( obj, res );
3409
3410 // Return final merged results
3411 set_control( _gvn.transform(region) );
3412 record_for_igvn(region);
3413
3414 bool not_null_free = !toop->can_be_value_type();
3415 bool not_flattenable = !ValueArrayFlatten || not_null_free || (toop->is_valuetypeptr() && !toop->value_klass()->flatten_array());
3416 if (EnableValhalla && not_flattenable) {
3417 // Check if obj has been loaded from an array
3418 obj = obj->isa_DecodeN() ? obj->in(1) : obj;
3419 Node* array = NULL;
3420 if (obj->isa_Load()) {
3421 Node* address = obj->in(MemNode::Address);
3422 if (address->isa_AddP()) {
3423 array = address->as_AddP()->in(AddPNode::Base);
3424 }
3425 } else if (obj->is_Phi()) {
3426 Node* region = obj->in(0);
3427 if (region->req() == 3 && region->in(2) != NULL) {
3428 IfNode* iff = region->in(2)->in(0)->isa_If();
3429 if (iff != NULL) {
3430 iff->is_flattened_array_check(&_gvn, array);
3431 }
3432 }
3433 }
3434 if (array != NULL) {
3435 const TypeAryPtr* ary_t = _gvn.type(array)->isa_aryptr();
3436 if (ary_t != NULL) {
3437 if (!ary_t->is_not_null_free() && not_null_free) {
3438 // Casting array element to a non-inline-type, mark array as not null-free.
3439 Node* cast = _gvn.transform(new CheckCastPPNode(control(), array, ary_t->cast_to_not_null_free()));
3440 replace_in_map(array, cast);
3441 } else if (!ary_t->is_not_flat()) {
3442 // Casting array element to a non-flattenable type, mark array as not flat.
3443 Node* cast = _gvn.transform(new CheckCastPPNode(control(), array, ary_t->cast_to_not_flat()));
3444 replace_in_map(array, cast);
3445 }
3446 }
3447 }
3448 }
3449
3450 if (!is_value) {
3451 res = record_profiled_receiver_for_speculation(res);
3452 if (toop->is_valuetypeptr() && toop->value_klass()->is_scalarizable() && !gvn().type(res)->maybe_null()) {
3453 res = ValueTypeNode::make_from_oop(this, res, toop->value_klass());
3454 }
3455 }
3456 return res;
3457 }
3458
3459 Node* GraphKit::is_always_locked(Node* obj) {
3460 Node* mark_addr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes());
3461 Node* mark = make_load(NULL, mark_addr, TypeX_X, TypeX_X->basic_type(), MemNode::unordered);
3462 Node* value_mask = _gvn.MakeConX(markOopDesc::always_locked_pattern);
3463 return _gvn.transform(new AndXNode(mark, value_mask));
3464 }
3465
3466 Node* GraphKit::is_value_mirror(Node* mirror) {
3467 Node* p = basic_plus_adr(mirror, java_lang_Class::inline_mirror_offset_in_bytes());
3468 Node* inline_mirror = access_load_at(mirror, p, _gvn.type(p)->is_ptr(), TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR), T_OBJECT, IN_HEAP);
3469 Node* cmp = _gvn.transform(new CmpPNode(mirror, inline_mirror));
3470 return _gvn.transform(new BoolNode(cmp, BoolTest::eq));
3471 }
3472
3473 // Deoptimize if 'obj' is a value type
3474 void GraphKit::gen_value_type_guard(Node* obj, int nargs) {
3475 assert(EnableValhalla, "should only be used if value types are enabled");
3476 Node* bol = NULL;
3477 if (obj->is_ValueTypeBase()) {
3478 bol = intcon(0);
3479 } else {
3480 Node* is_value = is_always_locked(obj);
3481 Node* value_mask = _gvn.MakeConX(markOopDesc::always_locked_pattern);
3482 Node* cmp = _gvn.transform(new CmpXNode(is_value, value_mask));
3483 bol = _gvn.transform(new BoolNode(cmp, BoolTest::ne));
3484 }
3485 { BuildCutout unless(this, bol, PROB_MAX);
3486 inc_sp(nargs);
3487 uncommon_trap(Deoptimization::Reason_class_check,
3488 Deoptimization::Action_none);
3489 }
3490 }
3491
3492 // Check if 'ary' is a null-free value type array
3493 Node* GraphKit::gen_null_free_array_check(Node* ary) {
3494 assert(EnableValhalla, "should only be used if value types are enabled");
3495 // Extract null free property from klass pointer
3496 Node* k_adr = basic_plus_adr(ary, oopDesc::klass_offset_in_bytes());
3497 const TypePtr* k_adr_type = k_adr->bottom_type()->isa_ptr();
3498 Node* klass = NULL;
3499 if (k_adr_type->is_ptr_to_narrowklass()) {
3500 klass = _gvn.transform(new LoadNKlassNode(NULL, immutable_memory(), k_adr, TypeInstPtr::KLASS, TypeKlassPtr::OBJECT->make_narrowklass(), MemNode::unordered));
3501 } else {
3502 klass = _gvn.transform(new LoadKlassNode(NULL, immutable_memory(), k_adr, TypeInstPtr::KLASS, TypeKlassPtr::OBJECT, MemNode::unordered));
3503 }
3504 Node* null_free = _gvn.transform(new GetNullFreePropertyNode(klass));
3505 Node* cmp = NULL;
3506 if (_gvn.type(klass)->isa_klassptr()) {
3507 cmp = _gvn.transform(new CmpLNode(null_free, zerocon(T_LONG)));
3508 } else {
3509 cmp = _gvn.transform(new CmpINode(null_free, zerocon(T_INT)));
3510 }
3511 return _gvn.transform(new BoolNode(cmp, BoolTest::eq));
3512 }
3513
3514 // Deoptimize if 'ary' is a null-free value type array and 'val' is null
3515 Node* GraphKit::gen_value_array_null_guard(Node* ary, Node* val, int nargs, bool safe_for_replace) {
3516 const Type* val_t = _gvn.type(val);
3517 if (val->is_ValueType() || !TypePtr::NULL_PTR->higher_equal(val_t)) {
3518 return ary; // Never null
3519 }
3520 RegionNode* region = new RegionNode(3);
3521 Node* null_ctl = top();
3522 null_check_oop(val, &null_ctl);
3523 if (null_ctl != top()) {
3524 PreserveJVMState pjvms(this);
3525 set_control(null_ctl);
3526 // Deoptimize if null-free array
3527 Node* bol = gen_null_free_array_check(ary);
3528 { BuildCutout unless(this, bol, PROB_MAX);
3529 inc_sp(nargs);
3530 uncommon_trap(Deoptimization::Reason_null_check,
3531 Deoptimization::Action_none);
3532 }
3533 region->init_req(1, control());
3534 }
3535 region->init_req(2, control());
3536 set_control(_gvn.transform(region));
3537 record_for_igvn(region);
3538 const TypeAryPtr* ary_t = _gvn.type(ary)->is_aryptr();
3539 if (val_t == TypePtr::NULL_PTR && !ary_t->is_not_null_free()) {
3540 // Since we were just successfully storing null, the array can't be null free.
3541 ary_t = ary_t->cast_to_not_null_free();
3542 Node* cast = _gvn.transform(new CheckCastPPNode(control(), ary, ary_t));
3543 if (safe_for_replace) {
3544 replace_in_map(ary, cast);
3545 }
3546 ary = cast;
3547 }
3548 return ary;
3549 }
3550
3551 Node* GraphKit::load_lh_array_tag(Node* kls) {
3552 Node* lhp = basic_plus_adr(kls, in_bytes(Klass::layout_helper_offset()));
3553 Node* layout_val = _gvn.transform(LoadNode::make(_gvn, NULL, immutable_memory(), lhp, lhp->bottom_type()->is_ptr(), TypeInt::INT, T_INT, MemNode::unordered));
3554
3555 return _gvn.transform(new RShiftINode(layout_val, intcon(Klass::_lh_array_tag_shift)));
3556 }
3557
3558
3559 Node* GraphKit::gen_lh_array_test(Node* kls, unsigned int lh_value) {
3560 Node* layout_val = load_lh_array_tag(kls);
3561 Node* cmp = _gvn.transform(new CmpINode(layout_val, intcon(lh_value)));
3562 return cmp;
3563 }
3564
3565
3566 //------------------------------next_monitor-----------------------------------
3567 // What number should be given to the next monitor?
3568 int GraphKit::next_monitor() {
3569 int current = jvms()->monitor_depth()* C->sync_stack_slots();
3570 int next = current + C->sync_stack_slots();
3571 // Keep the toplevel high water mark current:
3572 if (C->fixed_slots() < next) C->set_fixed_slots(next);
3573 return current;
3574 }
3575
3576 //------------------------------insert_mem_bar---------------------------------
3577 // Memory barrier to avoid floating things around
3578 // The membar serves as a pinch point between both control and all memory slices.
3579 Node* GraphKit::insert_mem_bar(int opcode, Node* precedent) {
3580 MemBarNode* mb = MemBarNode::make(C, opcode, Compile::AliasIdxBot, precedent);
3581 mb->init_req(TypeFunc::Control, control());
3582 mb->init_req(TypeFunc::Memory, reset_memory());
3583 Node* membar = _gvn.transform(mb);
3584 set_control(_gvn.transform(new ProjNode(membar, TypeFunc::Control)));
3585 set_all_memory_call(membar);
3623 Node* mem = reset_memory();
3624 MemBarNode* mb = MemBarNode::make(C, Op_MemBarVolatile, Compile::AliasIdxRaw);
3625 mb->init_req(TypeFunc::Control, control());
3626 mb->init_req(TypeFunc::Memory, mem);
3627 Node* membar = _gvn.transform(mb);
3628 set_control(_gvn.transform(new ProjNode(membar, TypeFunc::Control)));
3629 Node* newmem = _gvn.transform(new ProjNode(membar, TypeFunc::Memory));
3630 set_all_memory(mem);
3631 set_memory(newmem, Compile::AliasIdxRaw);
3632 }
3633
3634 //------------------------------shared_lock------------------------------------
3635 // Emit locking code.
3636 FastLockNode* GraphKit::shared_lock(Node* obj) {
3637 // bci is either a monitorenter bc or InvocationEntryBci
3638 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces
3639 assert(SynchronizationEntryBCI == InvocationEntryBci, "");
3640
3641 if( !GenerateSynchronizationCode )
3642 return NULL; // Not locking things?
3643
3644 if (stopped()) // Dead monitor?
3645 return NULL;
3646
3647 assert(dead_locals_are_killed(), "should kill locals before sync. point");
3648
3649 obj = access_resolve(obj, ACCESS_READ | ACCESS_WRITE);
3650
3651 // Box the stack location
3652 Node* box = _gvn.transform(new BoxLockNode(next_monitor()));
3653 Node* mem = reset_memory();
3654
3655 FastLockNode * flock = _gvn.transform(new FastLockNode(0, obj, box) )->as_FastLock();
3656 if (UseBiasedLocking && PrintPreciseBiasedLockingStatistics) {
3657 // Create the counters for this fast lock.
3658 flock->create_lock_counter(sync_jvms()); // sync_jvms used to get current bci
3659 }
3660
3661 // Create the rtm counters for this fast lock if needed.
3662 flock->create_rtm_lock_counter(sync_jvms()); // sync_jvms used to get current bci
3663
3698 }
3699 #endif
3700
3701 return flock;
3702 }
3703
3704
3705 //------------------------------shared_unlock----------------------------------
3706 // Emit unlocking code.
3707 void GraphKit::shared_unlock(Node* box, Node* obj) {
3708 // bci is either a monitorenter bc or InvocationEntryBci
3709 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces
3710 assert(SynchronizationEntryBCI == InvocationEntryBci, "");
3711
3712 if( !GenerateSynchronizationCode )
3713 return;
3714 if (stopped()) { // Dead monitor?
3715 map()->pop_monitor(); // Kill monitor from debug info
3716 return;
3717 }
3718 assert(!obj->is_ValueTypeBase(), "should not unlock on value type");
3719
3720 // Memory barrier to avoid floating things down past the locked region
3721 insert_mem_bar(Op_MemBarReleaseLock);
3722
3723 const TypeFunc *tf = OptoRuntime::complete_monitor_exit_Type();
3724 UnlockNode *unlock = new UnlockNode(C, tf);
3725 #ifdef ASSERT
3726 unlock->set_dbg_jvms(sync_jvms());
3727 #endif
3728 uint raw_idx = Compile::AliasIdxRaw;
3729 unlock->init_req( TypeFunc::Control, control() );
3730 unlock->init_req( TypeFunc::Memory , memory(raw_idx) );
3731 unlock->init_req( TypeFunc::I_O , top() ) ; // does no i/o
3732 unlock->init_req( TypeFunc::FramePtr, frameptr() );
3733 unlock->init_req( TypeFunc::ReturnAdr, top() );
3734
3735 unlock->init_req(TypeFunc::Parms + 0, obj);
3736 unlock->init_req(TypeFunc::Parms + 1, box);
3737 unlock = _gvn.transform(unlock)->as_Unlock();
3738
3739 Node* mem = reset_memory();
3740
3741 // unlock has no side-effects, sets few values
3742 set_predefined_output_for_runtime_call(unlock, mem, TypeRawPtr::BOTTOM);
3743
3744 // Kill monitor from debug info
3745 map()->pop_monitor( );
3746 }
3747
3748 //-------------------------------get_layout_helper-----------------------------
3749 // If the given klass is a constant or known to be an array,
3750 // fetch the constant layout helper value into constant_value
3751 // and return (Node*)NULL. Otherwise, load the non-constant
3752 // layout helper value, and return the node which represents it.
3753 // This two-faced routine is useful because allocation sites
3754 // almost always feature constant types.
3755 Node* GraphKit::get_layout_helper(Node* klass_node, jint& constant_value) {
3756 const TypeKlassPtr* inst_klass = _gvn.type(klass_node)->isa_klassptr();
3757 if (!StressReflectiveCode && inst_klass != NULL) {
3758 ciKlass* klass = inst_klass->klass();
3759 assert(klass != NULL, "klass should not be NULL");
3760 bool xklass = inst_klass->klass_is_exact();
3761 bool can_be_flattened = false;
3762 if (ValueArrayFlatten && klass->is_obj_array_klass()) {
3763 ciKlass* elem = klass->as_obj_array_klass()->element_klass();
3764 can_be_flattened = elem->is_java_lang_Object() || elem->is_interface() || (elem->is_valuetype() && !klass->as_array_klass()->storage_properties().is_null_free());
3765 }
3766 if (xklass || (klass->is_array_klass() && !can_be_flattened)) {
3767 jint lhelper = klass->layout_helper();
3768 if (lhelper != Klass::_lh_neutral_value) {
3769 constant_value = lhelper;
3770 return (Node*) NULL;
3771 }
3772 }
3773 }
3774 constant_value = Klass::_lh_neutral_value; // put in a known value
3775 Node* lhp = basic_plus_adr(klass_node, klass_node, in_bytes(Klass::layout_helper_offset()));
3776 return make_load(NULL, lhp, TypeInt::INT, T_INT, MemNode::unordered);
3777 }
3778
3779 // We just put in an allocate/initialize with a big raw-memory effect.
3780 // Hook selected additional alias categories on the initialization.
3781 static void hook_memory_on_init(GraphKit& kit, int alias_idx,
3782 MergeMemNode* init_in_merge,
3783 Node* init_out_raw) {
3784 DEBUG_ONLY(Node* init_in_raw = init_in_merge->base_memory());
3785 assert(init_in_merge->memory_at(alias_idx) == init_in_raw, "");
3786
3808
3809 // a normal slow-call doesn't change i_o, but an allocation does
3810 // we create a separate i_o projection for the normal control path
3811 set_i_o(_gvn.transform( new ProjNode(allocx, TypeFunc::I_O, false) ) );
3812 Node* rawoop = _gvn.transform( new ProjNode(allocx, TypeFunc::Parms) );
3813
3814 // put in an initialization barrier
3815 InitializeNode* init = insert_mem_bar_volatile(Op_Initialize, rawidx,
3816 rawoop)->as_Initialize();
3817 assert(alloc->initialization() == init, "2-way macro link must work");
3818 assert(init ->allocation() == alloc, "2-way macro link must work");
3819 {
3820 // Extract memory strands which may participate in the new object's
3821 // initialization, and source them from the new InitializeNode.
3822 // This will allow us to observe initializations when they occur,
3823 // and link them properly (as a group) to the InitializeNode.
3824 assert(init->in(InitializeNode::Memory) == malloc, "");
3825 MergeMemNode* minit_in = MergeMemNode::make(malloc);
3826 init->set_req(InitializeNode::Memory, minit_in);
3827 record_for_igvn(minit_in); // fold it up later, if possible
3828 _gvn.set_type(minit_in, Type::MEMORY);
3829 Node* minit_out = memory(rawidx);
3830 assert(minit_out->is_Proj() && minit_out->in(0) == init, "");
3831 // Add an edge in the MergeMem for the header fields so an access
3832 // to one of those has correct memory state
3833 set_memory(minit_out, C->get_alias_index(oop_type->add_offset(oopDesc::mark_offset_in_bytes())));
3834 set_memory(minit_out, C->get_alias_index(oop_type->add_offset(oopDesc::klass_offset_in_bytes())));
3835 if (oop_type->isa_aryptr()) {
3836 const TypeAryPtr* arytype = oop_type->is_aryptr();
3837 if (arytype->klass()->is_value_array_klass()) {
3838 // Initially all flattened array accesses share a single slice
3839 // but that changes after parsing. Prepare the memory graph so
3840 // it can optimize flattened array accesses properly once they
3841 // don't share a single slice.
3842 assert(C->flattened_accesses_share_alias(), "should be set at parse time");
3843 C->set_flattened_accesses_share_alias(false);
3844 ciValueArrayKlass* vak = arytype->klass()->as_value_array_klass();
3845 ciValueKlass* vk = vak->element_klass()->as_value_klass();
3846 for (int i = 0, len = vk->nof_nonstatic_fields(); i < len; i++) {
3847 ciField* field = vk->nonstatic_field_at(i);
3848 if (field->offset() >= TrackedInitializationLimit * HeapWordSize)
3849 continue; // do not bother to track really large numbers of fields
3850 int off_in_vt = field->offset() - vk->first_field_offset();
3851 const TypePtr* adr_type = arytype->with_field_offset(off_in_vt)->add_offset(Type::OffsetBot);
3852 int fieldidx = C->get_alias_index(adr_type, true);
3853 hook_memory_on_init(*this, fieldidx, minit_in, minit_out);
3854 }
3855 C->set_flattened_accesses_share_alias(true);
3856 hook_memory_on_init(*this, C->get_alias_index(TypeAryPtr::VALUES), minit_in, minit_out);
3857 } else {
3858 const TypePtr* telemref = oop_type->add_offset(Type::OffsetBot);
3859 int elemidx = C->get_alias_index(telemref);
3860 hook_memory_on_init(*this, elemidx, minit_in, minit_out);
3861 }
3862 } else if (oop_type->isa_instptr()) {
3863 set_memory(minit_out, C->get_alias_index(oop_type)); // mark word
3864 ciInstanceKlass* ik = oop_type->klass()->as_instance_klass();
3865 for (int i = 0, len = ik->nof_nonstatic_fields(); i < len; i++) {
3866 ciField* field = ik->nonstatic_field_at(i);
3867 if (field->offset() >= TrackedInitializationLimit * HeapWordSize)
3868 continue; // do not bother to track really large numbers of fields
3869 // Find (or create) the alias category for this field:
3870 int fieldidx = C->alias_type(field)->index();
3871 hook_memory_on_init(*this, fieldidx, minit_in, minit_out);
3872 }
3873 }
3874 }
3875
3876 // Cast raw oop to the real thing...
3877 Node* javaoop = new CheckCastPPNode(control(), rawoop, oop_type);
3878 javaoop = _gvn.transform(javaoop);
3879 C->set_recent_alloc(control(), javaoop);
3880 assert(just_allocated_object(control()) == javaoop, "just allocated");
3881
3882 #ifdef ASSERT
3883 { // Verify that the AllocateNode::Ideal_allocation recognizers work:
3894 assert(alloc->in(AllocateNode::ALength)->is_top(), "no length, please");
3895 }
3896 }
3897 #endif //ASSERT
3898
3899 return javaoop;
3900 }
3901
3902 //---------------------------new_instance--------------------------------------
3903 // This routine takes a klass_node which may be constant (for a static type)
3904 // or may be non-constant (for reflective code). It will work equally well
3905 // for either, and the graph will fold nicely if the optimizer later reduces
3906 // the type to a constant.
3907 // The optional arguments are for specialized use by intrinsics:
3908 // - If 'extra_slow_test' if not null is an extra condition for the slow-path.
3909 // - If 'return_size_val', report the the total object size to the caller.
3910 // - deoptimize_on_exception controls how Java exceptions are handled (rethrow vs deoptimize)
3911 Node* GraphKit::new_instance(Node* klass_node,
3912 Node* extra_slow_test,
3913 Node* *return_size_val,
3914 bool deoptimize_on_exception,
3915 ValueTypeBaseNode* value_node) {
3916 // Compute size in doublewords
3917 // The size is always an integral number of doublewords, represented
3918 // as a positive bytewise size stored in the klass's layout_helper.
3919 // The layout_helper also encodes (in a low bit) the need for a slow path.
3920 jint layout_con = Klass::_lh_neutral_value;
3921 Node* layout_val = get_layout_helper(klass_node, layout_con);
3922 bool layout_is_con = (layout_val == NULL);
3923
3924 if (extra_slow_test == NULL) extra_slow_test = intcon(0);
3925 // Generate the initial go-slow test. It's either ALWAYS (return a
3926 // Node for 1) or NEVER (return a NULL) or perhaps (in the reflective
3927 // case) a computed value derived from the layout_helper.
3928 Node* initial_slow_test = NULL;
3929 if (layout_is_con) {
3930 assert(!StressReflectiveCode, "stress mode does not use these paths");
3931 bool must_go_slow = Klass::layout_helper_needs_slow_path(layout_con);
3932 initial_slow_test = must_go_slow ? intcon(1) : extra_slow_test;
3933 } else { // reflective case
3934 // This reflective path is used by Unsafe.allocateInstance.
3935 // (It may be stress-tested by specifying StressReflectiveCode.)
3936 // Basically, we want to get into the VM is there's an illegal argument.
3937 Node* bit = intcon(Klass::_lh_instance_slow_path_bit);
3938 initial_slow_test = _gvn.transform( new AndINode(layout_val, bit) );
3939 if (extra_slow_test != intcon(0)) {
3940 initial_slow_test = _gvn.transform( new OrINode(initial_slow_test, extra_slow_test) );
3941 }
3942 // (Macro-expander will further convert this to a Bool, if necessary.)
3953
3954 // Clear the low bits to extract layout_helper_size_in_bytes:
3955 assert((int)Klass::_lh_instance_slow_path_bit < BytesPerLong, "clear bit");
3956 Node* mask = MakeConX(~ (intptr_t)right_n_bits(LogBytesPerLong));
3957 size = _gvn.transform( new AndXNode(size, mask) );
3958 }
3959 if (return_size_val != NULL) {
3960 (*return_size_val) = size;
3961 }
3962
3963 // This is a precise notnull oop of the klass.
3964 // (Actually, it need not be precise if this is a reflective allocation.)
3965 // It's what we cast the result to.
3966 const TypeKlassPtr* tklass = _gvn.type(klass_node)->isa_klassptr();
3967 if (!tklass) tklass = TypeKlassPtr::OBJECT;
3968 const TypeOopPtr* oop_type = tklass->as_instance_type();
3969
3970 // Now generate allocation code
3971
3972 // The entire memory state is needed for slow path of the allocation
3973 // since GC and deoptimization can happen.
3974 Node *mem = reset_memory();
3975 set_all_memory(mem); // Create new memory state
3976
3977 AllocateNode* alloc = new AllocateNode(C, AllocateNode::alloc_type(Type::TOP),
3978 control(), mem, i_o(),
3979 size, klass_node,
3980 initial_slow_test, value_node);
3981
3982 return set_output_for_allocation(alloc, oop_type, deoptimize_on_exception);
3983 }
3984
3985 // With compressed oops, the 64 bit init value for non flattened value
3986 // arrays is built from 2 32 bit compressed oops
3987 static Node* raw_default_for_coops(Node* default_value, GraphKit& kit) {
3988 Node* lower = kit.gvn().transform(new CastP2XNode(kit.control(), default_value));
3989 Node* upper = kit.gvn().transform(new LShiftLNode(lower, kit.intcon(32)));
3990 return kit.gvn().transform(new OrLNode(lower, upper));
3991 }
3992
3993 //-------------------------------new_array-------------------------------------
3994 // helper for newarray and anewarray
3995 // The 'length' parameter is (obviously) the length of the array.
3996 // See comments on new_instance for the meaning of the other arguments.
3997 Node* GraphKit::new_array(Node* klass_node, // array klass (maybe variable)
3998 Node* length, // number of array elements
3999 int nargs, // number of arguments to push back for uncommon trap
4000 Node* *return_size_val,
4001 bool deoptimize_on_exception,
4002 Node* elem_mirror) {
4003 jint layout_con = Klass::_lh_neutral_value;
4004 Node* layout_val = get_layout_helper(klass_node, layout_con);
4005 bool layout_is_con = (layout_val == NULL);
4006
4007 if (!layout_is_con && !StressReflectiveCode &&
4008 !too_many_traps(Deoptimization::Reason_class_check)) {
4009 // This is a reflective array creation site.
4010 // Optimistically assume that it is a subtype of Object[],
4011 // so that we can fold up all the address arithmetic.
4012 layout_con = Klass::array_layout_helper(T_OBJECT);
4013 Node* cmp_lh = _gvn.transform( new CmpINode(layout_val, intcon(layout_con)) );
4014 Node* bol_lh = _gvn.transform( new BoolNode(cmp_lh, BoolTest::eq) );
4015 { BuildCutout unless(this, bol_lh, PROB_MAX);
4016 inc_sp(nargs);
4017 uncommon_trap(Deoptimization::Reason_class_check,
4018 Deoptimization::Action_maybe_recompile);
4019 }
4020 layout_val = NULL;
4021 layout_is_con = true;
4022 }
4023
4024 // Generate the initial go-slow test. Make sure we do not overflow
4025 // if length is huge (near 2Gig) or negative! We do not need
4026 // exact double-words here, just a close approximation of needed
4027 // double-words. We can't add any offset or rounding bits, lest we
4028 // take a size -1 of bytes and make it positive. Use an unsigned
4029 // compare, so negative sizes look hugely positive.
4030 int fast_size_limit = FastAllocateSizeLimit;
4031 if (layout_is_con) {
4032 assert(!StressReflectiveCode, "stress mode does not use these paths");
4033 // Increase the size limit if we have exact knowledge of array type.
4034 int log2_esize = Klass::layout_helper_log2_element_size(layout_con);
4035 fast_size_limit <<= MAX2(LogBytesPerLong - log2_esize, 0);
4036 }
4037
4038 Node* initial_slow_cmp = _gvn.transform( new CmpUNode( length, intcon( fast_size_limit ) ) );
4039 Node* initial_slow_test = _gvn.transform( new BoolNode( initial_slow_cmp, BoolTest::gt ) );
4040
4041 // --- Size Computation ---
4042 // array_size = round_to_heap(array_header + (length << elem_shift));
4043 // where round_to_heap(x) == align_to(x, MinObjAlignmentInBytes)
4044 // and align_to(x, y) == ((x + y-1) & ~(y-1))
4045 // The rounding mask is strength-reduced, if possible.
4046 int round_mask = MinObjAlignmentInBytes - 1;
4047 Node* header_size = NULL;
4048 int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE);
4049 // (T_BYTE has the weakest alignment and size restrictions...)
4050 if (layout_is_con) {
4051 int hsize = Klass::layout_helper_header_size(layout_con);
4052 int eshift = Klass::layout_helper_log2_element_size(layout_con);
4053 bool is_value_array = Klass::layout_helper_is_valueArray(layout_con);
4054 if ((round_mask & ~right_n_bits(eshift)) == 0)
4055 round_mask = 0; // strength-reduce it if it goes away completely
4056 assert(is_value_array || (hsize & right_n_bits(eshift)) == 0, "hsize is pre-rounded");
4057 assert(header_size_min <= hsize, "generic minimum is smallest");
4058 header_size_min = hsize;
4059 header_size = intcon(hsize + round_mask);
4060 } else {
4061 Node* hss = intcon(Klass::_lh_header_size_shift);
4062 Node* hsm = intcon(Klass::_lh_header_size_mask);
4063 Node* hsize = _gvn.transform( new URShiftINode(layout_val, hss) );
4064 hsize = _gvn.transform( new AndINode(hsize, hsm) );
4065 Node* mask = intcon(round_mask);
4066 header_size = _gvn.transform( new AddINode(hsize, mask) );
4067 }
4068
4069 Node* elem_shift = NULL;
4070 if (layout_is_con) {
4071 int eshift = Klass::layout_helper_log2_element_size(layout_con);
4072 if (eshift != 0)
4073 elem_shift = intcon(eshift);
4074 } else {
4075 // There is no need to mask or shift this value.
4076 // The semantics of LShiftINode include an implicit mask to 0x1F.
4120 // places, one where the length is sharply limited, and the other
4121 // after a successful allocation.
4122 Node* abody = lengthx;
4123 if (elem_shift != NULL)
4124 abody = _gvn.transform( new LShiftXNode(lengthx, elem_shift) );
4125 Node* size = _gvn.transform( new AddXNode(headerx, abody) );
4126 if (round_mask != 0) {
4127 Node* mask = MakeConX(~round_mask);
4128 size = _gvn.transform( new AndXNode(size, mask) );
4129 }
4130 // else if round_mask == 0, the size computation is self-rounding
4131
4132 if (return_size_val != NULL) {
4133 // This is the size
4134 (*return_size_val) = size;
4135 }
4136
4137 // Now generate allocation code
4138
4139 // The entire memory state is needed for slow path of the allocation
4140 // since GC and deoptimization can happen.
4141 Node *mem = reset_memory();
4142 set_all_memory(mem); // Create new memory state
4143
4144 if (initial_slow_test->is_Bool()) {
4145 // Hide it behind a CMoveI, or else PhaseIdealLoop::split_up will get sick.
4146 initial_slow_test = initial_slow_test->as_Bool()->as_int_value(&_gvn);
4147 }
4148
4149 const TypeOopPtr* ary_type = _gvn.type(klass_node)->is_klassptr()->as_instance_type();
4150 const TypeAryPtr* ary_ptr = ary_type->isa_aryptr();
4151 const Type* elem = NULL;
4152 ciKlass* elem_klass = NULL;
4153
4154 // Compute default value and storage properties for value type arrays:
4155 // - null-ok: MyValue.box[] (ciObjArrayKlass "[LMyValue")
4156 // - null-free: MyValue.val[] (ciObjArrayKlass "[QMyValue")
4157 // - null-free, flattened: MyValue.val[] (ciValueArrayKlass "[QMyValue")
4158 Node* storage_properties = NULL;
4159 Node* default_value = NULL;
4160 Node* raw_default_value = NULL;
4161 int props_shift = UseCompressedClassPointers ? oopDesc::narrow_storage_props_shift : oopDesc::wide_storage_props_shift;
4162 if (ary_ptr != NULL && ary_ptr->klass_is_exact()) {
4163 // Array type is known
4164 elem = ary_ptr->elem();
4165 ciArrayKlass* ary_klass = ary_ptr->klass()->as_array_klass();
4166 elem_klass = ary_klass->element_klass();
4167
4168 ArrayStorageProperties props = ary_klass->storage_properties();
4169 if (!props.is_empty() && elem_klass->is_valuetype()) {
4170 if (props.is_null_free() && !props.is_flattened()) {
4171 default_value = ValueTypeNode::default_oop(gvn(), elem_klass->as_value_klass());
4172 if (elem->isa_narrowoop()) {
4173 default_value = _gvn.transform(new EncodePNode(default_value, elem));
4174 raw_default_value = raw_default_for_coops(default_value, *this);
4175 } else {
4176 raw_default_value = _gvn.transform(new CastP2XNode(control(), default_value));
4177 }
4178 }
4179 storage_properties = MakeConX(props.encode<NOT_LP64(jint) LP64_ONLY(jlong)>(props_shift));
4180 }
4181 }
4182
4183 if (EnableValhalla && (elem == NULL || (elem_klass != NULL && (elem_klass->is_java_lang_Object() || elem_klass->is_valuetype()) &&
4184 !ary_type->klass_is_exact()))) {
4185 // Array type is not known, compute default value and storage properties for initialization.
4186 assert(default_value == NULL && raw_default_value == NULL && storage_properties == NULL, "shouldn't be set yet");
4187 assert(elem_mirror != NULL, "should not be null");
4188
4189 Node* r = new RegionNode(4);
4190 default_value = new PhiNode(r, TypeInstPtr::BOTTOM);
4191 storage_properties = new PhiNode(r, TypeX_X);
4192
4193 Node* empty = MakeConX(ArrayStorageProperties::empty.encode<NOT_LP64(jint) LP64_ONLY(jlong)>(props_shift));
4194 Node* null_free = MakeConX(ArrayStorageProperties::null_free.encode<NOT_LP64(jint) LP64_ONLY(jlong)>(props_shift));
4195 Node* flat = MakeConX(ArrayStorageProperties::flattened_and_null_free.encode<NOT_LP64(jint) LP64_ONLY(jlong)>(props_shift));
4196
4197 // Check if element mirror is a value mirror
4198 IfNode* iff = create_and_map_if(control(), is_value_mirror(elem_mirror), PROB_FAIR, COUNT_UNKNOWN);
4199
4200 // Not a value mirror but a box mirror or not a value type array, initialize with all zero
4201 r->init_req(1, _gvn.transform(new IfFalseNode(iff)));
4202 default_value->init_req(1, null());
4203 storage_properties->init_req(1, empty);
4204
4205 // Value mirror (= null-free), check if flattened
4206 set_control(_gvn.transform(new IfTrueNode(iff)));
4207 Node* cmp = gen_lh_array_test(klass_node, Klass::_lh_array_tag_vt_value);
4208 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::eq));
4209 iff = create_and_map_if(control(), bol, PROB_FAIR, COUNT_UNKNOWN);
4210
4211 // Flattened, initialize with all zero
4212 r->init_req(2, _gvn.transform(new IfTrueNode(iff)));
4213 default_value->init_req(2, null());
4214 storage_properties->init_req(2, flat);
4215
4216 // Non-flattened, initialize with the default value
4217 set_control(_gvn.transform(new IfFalseNode(iff)));
4218 Node* p = basic_plus_adr(klass_node, in_bytes(ArrayKlass::element_klass_offset()));
4219 Node* eklass = _gvn.transform(LoadKlassNode::make(_gvn, control(), immutable_memory(), p, TypeInstPtr::KLASS));
4220 Node* adr_fixed_block_addr = basic_plus_adr(eklass, in_bytes(InstanceKlass::adr_valueklass_fixed_block_offset()));
4221 Node* adr_fixed_block = make_load(control(), adr_fixed_block_addr, TypeRawPtr::NOTNULL, T_ADDRESS, MemNode::unordered);
4222 Node* default_value_offset_addr = basic_plus_adr(adr_fixed_block, in_bytes(ValueKlass::default_value_offset_offset()));
4223 Node* default_value_offset = make_load(control(), default_value_offset_addr, TypeInt::INT, T_INT, MemNode::unordered);
4224 Node* default_value_addr = basic_plus_adr(elem_mirror, ConvI2X(default_value_offset));
4225 Node* val = access_load_at(elem_mirror, default_value_addr, _gvn.type(default_value_addr)->is_ptr(), TypeInstPtr::BOTTOM, T_OBJECT, IN_HEAP);
4226 r->init_req(3, control());
4227 default_value->init_req(3, val);
4228 storage_properties->init_req(3, null_free);
4229
4230 set_control(_gvn.transform(r));
4231 default_value = _gvn.transform(default_value);
4232 storage_properties = _gvn.transform(storage_properties);
4233 if (UseCompressedOops) {
4234 default_value = _gvn.transform(new EncodePNode(default_value, default_value->bottom_type()->make_narrowoop()));
4235 raw_default_value = raw_default_for_coops(default_value, *this);
4236 } else {
4237 raw_default_value = _gvn.transform(new CastP2XNode(control(), default_value));
4238 }
4239 }
4240
4241 // Create the AllocateArrayNode and its result projections
4242 AllocateArrayNode* alloc = new AllocateArrayNode(C, AllocateArrayNode::alloc_type(TypeInt::INT),
4243 control(), mem, i_o(),
4244 size, klass_node,
4245 initial_slow_test,
4246 length, default_value,
4247 raw_default_value,
4248 storage_properties);
4249
4250 // Cast to correct type. Note that the klass_node may be constant or not,
4251 // and in the latter case the actual array type will be inexact also.
4252 // (This happens via a non-constant argument to inline_native_newArray.)
4253 // In any case, the value of klass_node provides the desired array type.
4254 const TypeInt* length_type = _gvn.find_int_type(length);
4255 if (ary_type->isa_aryptr() && length_type != NULL) {
4256 // Try to get a better type than POS for the size
4257 ary_type = ary_type->is_aryptr()->cast_to_size(length_type);
4258 }
4259
4260 Node* javaoop = set_output_for_allocation(alloc, ary_type, deoptimize_on_exception);
4261
4262 // Cast length on remaining path to be as narrow as possible
4263 if (map()->find_edge(length) >= 0) {
4264 Node* ccast = alloc->make_ideal_length(ary_type, &_gvn);
4265 if (ccast != length) {
4266 _gvn.set_type_bottom(ccast);
4267 record_for_igvn(ccast);
4268 replace_in_map(length, ccast);
4269 }
4270 }
4271
4272 return javaoop;
4273 }
4274
4389 set_all_memory(ideal.merged_memory());
4390 set_i_o(ideal.i_o());
4391 set_control(ideal.ctrl());
4392 }
4393
4394 void GraphKit::final_sync(IdealKit& ideal) {
4395 // Final sync IdealKit and graphKit.
4396 sync_kit(ideal);
4397 }
4398
4399 Node* GraphKit::load_String_length(Node* str, bool set_ctrl) {
4400 Node* len = load_array_length(load_String_value(str, set_ctrl));
4401 Node* coder = load_String_coder(str, set_ctrl);
4402 // Divide length by 2 if coder is UTF16
4403 return _gvn.transform(new RShiftINode(len, coder));
4404 }
4405
4406 Node* GraphKit::load_String_value(Node* str, bool set_ctrl) {
4407 int value_offset = java_lang_String::value_offset_in_bytes();
4408 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4409 false, NULL, Type::Offset(0));
4410 const TypePtr* value_field_type = string_type->add_offset(value_offset);
4411 const TypeAryPtr* value_type = TypeAryPtr::make(TypePtr::NotNull,
4412 TypeAry::make(TypeInt::BYTE, TypeInt::POS, false, true, true),
4413 ciTypeArrayKlass::make(T_BYTE), true, Type::Offset(0));
4414 Node* p = basic_plus_adr(str, str, value_offset);
4415 Node* load = access_load_at(str, p, value_field_type, value_type, T_OBJECT,
4416 IN_HEAP | (set_ctrl ? C2_CONTROL_DEPENDENT_LOAD : 0) | MO_UNORDERED);
4417 return load;
4418 }
4419
4420 Node* GraphKit::load_String_coder(Node* str, bool set_ctrl) {
4421 if (!CompactStrings) {
4422 return intcon(java_lang_String::CODER_UTF16);
4423 }
4424 int coder_offset = java_lang_String::coder_offset_in_bytes();
4425 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4426 false, NULL, Type::Offset(0));
4427 const TypePtr* coder_field_type = string_type->add_offset(coder_offset);
4428
4429 Node* p = basic_plus_adr(str, str, coder_offset);
4430 Node* load = access_load_at(str, p, coder_field_type, TypeInt::BYTE, T_BYTE,
4431 IN_HEAP | (set_ctrl ? C2_CONTROL_DEPENDENT_LOAD : 0) | MO_UNORDERED);
4432 return load;
4433 }
4434
4435 void GraphKit::store_String_value(Node* str, Node* value) {
4436 int value_offset = java_lang_String::value_offset_in_bytes();
4437 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4438 false, NULL, Type::Offset(0));
4439 const TypePtr* value_field_type = string_type->add_offset(value_offset);
4440
4441 access_store_at(str, basic_plus_adr(str, value_offset), value_field_type,
4442 value, TypeAryPtr::BYTES, T_OBJECT, IN_HEAP | MO_UNORDERED);
4443 }
4444
4445 void GraphKit::store_String_coder(Node* str, Node* value) {
4446 int coder_offset = java_lang_String::coder_offset_in_bytes();
4447 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4448 false, NULL, Type::Offset(0));
4449 const TypePtr* coder_field_type = string_type->add_offset(coder_offset);
4450
4451 access_store_at(str, basic_plus_adr(str, coder_offset), coder_field_type,
4452 value, TypeInt::BYTE, T_BYTE, IN_HEAP | MO_UNORDERED);
4453 }
4454
4455 // Capture src and dst memory state with a MergeMemNode
4456 Node* GraphKit::capture_memory(const TypePtr* src_type, const TypePtr* dst_type) {
4457 if (src_type == dst_type) {
4458 // Types are equal, we don't need a MergeMemNode
4459 return memory(src_type);
4460 }
4461 MergeMemNode* merge = MergeMemNode::make(map()->memory());
4462 record_for_igvn(merge); // fold it up later, if possible
4463 int src_idx = C->get_alias_index(src_type);
4464 int dst_idx = C->get_alias_index(dst_type);
4465 merge->set_memory_at(src_idx, memory(src_idx));
4466 merge->set_memory_at(dst_idx, memory(dst_idx));
4467 return merge;
4468 }
4541 i_char->init_req(2, AddI(i_char, intcon(2)));
4542
4543 set_control(IfFalse(iff));
4544 set_memory(st, TypeAryPtr::BYTES);
4545 }
4546
4547 Node* GraphKit::make_constant_from_field(ciField* field, Node* obj) {
4548 if (!field->is_constant()) {
4549 return NULL; // Field not marked as constant.
4550 }
4551 ciInstance* holder = NULL;
4552 if (!field->is_static()) {
4553 ciObject* const_oop = obj->bottom_type()->is_oopptr()->const_oop();
4554 if (const_oop != NULL && const_oop->is_instance()) {
4555 holder = const_oop->as_instance();
4556 }
4557 }
4558 const Type* con_type = Type::make_constant_from_field(field, holder, field->layout_type(),
4559 /*is_unsigned_load=*/false);
4560 if (con_type != NULL) {
4561 Node* con = makecon(con_type);
4562 if (field->layout_type() == T_VALUETYPE && field->type()->as_value_klass()->is_scalarizable() && !con_type->maybe_null()) {
4563 // Load value type from constant oop
4564 con = ValueTypeNode::make_from_oop(this, con, field->type()->as_value_klass());
4565 }
4566 return con;
4567 }
4568 return NULL;
4569 }
4570
4571 //---------------------------load_mirror_from_klass----------------------------
4572 // Given a klass oop, load its java mirror (a java.lang.Class oop).
4573 Node* GraphKit::load_mirror_from_klass(Node* klass) {
4574 Node* p = basic_plus_adr(klass, in_bytes(Klass::java_mirror_offset()));
4575 Node* load = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS, MemNode::unordered);
4576 // mirror = ((OopHandle)mirror)->resolve();
4577 return access_load(load, TypeInstPtr::MIRROR, T_OBJECT, IN_NATIVE);
4578 }
|